Numerical Solution of Ordinary Differential Equations Endre S¨ uli

Transcription

Numerical Solution of Ordinary Differential Equations Endre S¨ uli
Numerical Solution of Ordinary
Differential Equations
Endre S¨
uli
Mathematical Institute, University of Oxford
October 18, 2014
1
Contents
1 Picard’s theorem
1
2 One-step methods
2.1 Euler’s method and its relatives: the θ-method
2.2 Error analysis of the θ-method . . . . . . . . .
2.3 General explicit one-step method . . . . . . . .
2.4 Runge–Kutta methods . . . . . . . . . . . . . .
2.5 Absolute stability of Runge–Kutta methods . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4
4
7
9
13
19
3 Linear multi-step methods
3.1 Construction of linear multi-step methods . . . . . . . . . . .
3.2 Zero-stability . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3 Consistency . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.4 Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.4.1 Necessary conditions for convergence . . . . . . . . . .
3.4.2 Sufficient conditions for convergence . . . . . . . . . .
3.5 Maximum order of a zero-stable linear multi-step method . .
3.6 Absolute stability of linear multistep methods . . . . . . . . .
3.7 General methods for locating the interval of absolute stability
3.7.1 The Schur criterion . . . . . . . . . . . . . . . . . . . .
3.7.2 The Routh–Hurwitz criterion . . . . . . . . . . . . . .
3.8 Predictor-corrector methods . . . . . . . . . . . . . . . . . . .
3.8.1 Absolute stability of predictor-corrector methods . . .
3.8.2 The accuracy of predictor-corrector methods . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
21
22
24
26
29
30
33
37
43
45
45
46
48
50
52
4 Stiff problems
4.1 Stability of numerical methods for stiff systems . . . . . . . . . . . . . . . . . . . .
4.2 Backward differentiation methods for stiff systems . . . . . . . . . . . . . . . . . .
4.3 Gear’s method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
53
54
57
58
5 Nonlinear Stability
59
6 Boundary value problems
6.1 Shooting methods . . . . . . . . . . . . .
6.1.1 The method of bisection . . . . . .
6.1.2 The Newton–Raphson method . .
6.2 Matrix methods . . . . . . . . . . . . . . .
6.2.1 Linear boundary value problem . .
6.2.2 Nonlinear boundary value problem
6.3 Collocation method . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
62
62
63
63
66
66
69
70
Preface.
The purpose of these lecture notes is to provide an introduction to computational methods for the approximate solution of ordinary differential equations (ODEs).
Only minimal prerequisites in differential and integral calculus, differential equation theory, complex analysis and linear algebra are assumed. The notes focus on the construction
of numerical algorithms for ODEs and the mathematical analysis of their behaviour, covering the material taught in the M.Sc. in Mathematical Modelling and Scientific Computation in the eight-lecture course Numerical Solution of Ordinary Differential Equations.
The notes begin with a study of well-posedness of initial value problems for a firstorder differential equations and systems of such equations. The basic ideas of discretisation
and error analysis are then introduced in the case of one-step methods. This is followed
by an extension of the concepts of stability and accuracy to linear multi-step methods,
including predictor corrector methods, and a brief excursion into numerical methods for
stiff systems of ODEs. The final sections are devoted to an overview of classical algorithms
for the numerical solution of two-point boundary value problems.
Syllabus.
Approximation of initial value problems for ordinary differential equations:
one-step methods including the explicit and implicit Euler methods, the trapezium rule
method, and Runge–Kutta methods. Linear multi-step methods: consistency, zerostability and convergence; absolute stability. Predictor-corrector methods.
Stiffness, stability regions, Gear’s methods and their implementation. Nonlinear stability.
Boundary value problems: shooting methods, matrix methods and collocation.
Reading List:
[1] H.B. Keller, Numerical Methods for Two-point Boundary Value Problems. SIAM,
Philadelphia, 1976.
[2] J.D. Lambert, Computational Methods in Ordinary Differential Equations. Wiley,
Chichester, 1991.
Further Reading:
[1] E. Hairer, S.P. Norsett, and G. Wanner, Solving Ordinary Differential Equations I: Nonstiff Problems. Springer-Verlag, Berlin, 1987.
[2] P. Henrici, Discrete Variable Methods in Ordinary Differential Equations. Wiley,
New York, 1962.
[3] K.W. Morton, Numerical Solution of Ordinary Differential Equations. Oxford
University Computing Laboratory, 1987.
[4] A.M. Stuart and A.R. Humphries, Dynamical Systems and Numerical Analysis.
Cambridge University Press, Cambridge, 1996.
Acknowledgement: I am grateful to Ms M. H. Sadeghian (Department Of Mathematics, Ferdowsi University of Mashhad) for numerous helpful comments and corrections.
1
Picard’s theorem
Ordinary differential equations frequently occur as mathematical models in many branches
of science, engineering and economy. Unfortunately it is seldom that these equations have
solutions that can be expressed in closed form, so it is common to seek approximate
solutions by means of numerical methods; nowadays this can usually be achieved very inexpensively to high accuracy and with a reliable bound on the error between the analytical
solution and its numerical approximation. In this section we shall be concerned with the
construction and the analysis of numerical methods for first-order differential equations of
the form
y ′ = f (x, y)
(1)
for the real-valued function y of the real variable x, where y ′ ≡ dy/dx. In order to select
a particular integral from the infinite family of solution curves that constitute the general
solution to (1), the differential equation will be considered in tandem with an initial
condition: given two real numbers x0 and y0 , we seek a solution to (1) for x > x0 such
that
(2)
y(x0 ) = y0 .
The differential equation (1) together with the initial condition (2) is called an initial
value problem.
In general, even if f (·, ·) is a continuous function, there is no guarantee that the
initial value problem (1–2) possesses a unique solution1 . Fortunately, under a further mild
condition on the function f , the existence and uniqueness of a solution to (1–2) can be
ensured: the result is encapsulated in the next theorem.
Theorem 1 (Picard’s Theorem2 ) Suppose that f (·, ·) is a continuous function of its
arguments in a region U of the (x, y) plane which contains the rectangle
R = {(x, y) : x0 ≤ x ≤ XM ,
|y − y0 | ≤ YM } ,
where XM > x0 and YM > 0 are constants. Suppose also, that there exists a positive
constant L such that
|f (x, y) − f (x, z)| ≤ L|y − z|
(3)
holds whenever (x, y) and (x, z) lie in the rectangle R. Finally, letting
M = max{|f (x, y)| : (x, y) ∈ R} ,
suppose that M (XM − x0 ) ≤ YM . Then there exists a unique continuously differentiable
function x 7→ y(x), defined on the closed interval [x0 , XM ], which satisfies (1) and (2).
The condition (3) is called a Lipschitz condition3 , and L is called the Lipschitz
constant for f .
We shall not dwell on the proof of Picard’s Theorem; for details, the interested reader
is referred to any good textbook on the theory of ordinary differential equations, or the
1
Consider, for example, the initial value problem y ′ = y 2/3 , y(0) = 0; this has two solutions: y(x) ≡ 0
and y(x) = x3 /27.
2
Emile Picard (1856–1941)
3
Rudolf Lipschitz (1832–1903)
1
lecture notes of P. J. Collins, Differential and Integral Equations, Part I, Mathematical Institute Oxford, 1988 (reprinted 1990). The essence of the proof is to consider the sequence
of functions {yn }∞
n=0 , defined recursively through what is known as the Picard Iteration:
y0 (x) ≡ y0 ,
yn (x) = y0 +
(4)
Z
x
x0
f (ξ, yn−1 (ξ)) dξ ,
n = 1, 2, . . . ,
and show, using the conditions of the theorem, that {yn }∞
n=0 converges uniformly on the
interval [x0 , XM ] to a function y defined on [x0 , XM ] such that
Z
y(x) = y0 +
x
f (ξ, y(ξ)) dξ .
x0
This then implies that y is continuously differentiable on [x0 , XM ] and it satisfies the
differential equation (1) and the initial condition (2). The uniqueness of the solution
follows from the Lipschitz condition.
Picard’s Theorem has a natural extension to an initial value problem for a system of
m differential equations of the form
y′ = f (x, y) ,
y(x0 ) = y0 ,
(5)
where y0 ∈ Rm and f : [x0 , XM ] × Rm → Rm . On introducing the Euclidean norm k · k
on Rm by
kvk =
we can state the following result.
m
X
i=1
2
|vi |
!1/2
,
v ∈ Rm ,
Theorem 2 (Picard’s Theorem) Suppose that f (·, ·) is a continuous function of its
arguments in a region U of the (x, y) space R1+m which contains the parallelepiped
R = {(x, y) : x0 ≤ x ≤ XM ,
ky − y0 k ≤ YM } ,
where XM > x0 and YM > 0 are constants. Suppose also that there exists a positive
constant L such that
kf (x, y) − f (x, z)k ≤ Lky − zk
(6)
holds whenever (x, y) and (x, z) lie in R. Finally, letting
M = max{kf (x, y)k : (x, y) ∈ R} ,
suppose that M (XM − x0 ) ≤ YM . Then there exists a unique continuously differentiable
function x 7→ y(x), defined on the closed interval [x0 , XM ], which satisfies (5).
A sufficient condition for (6) is that f is continuous on R, differentiable at each point
(x, y) in int(R), the interior of R, and there exists L > 0 such that
∂f
≤ L
(x,
y)
∂y
for all (x, y) ∈ int(R) ,
2
(7)
where ∂f /∂y denotes the m × m Jacobi matrix of y ∈ Rm 7→ f (x, y) ∈ Rm , and k · k is a
matrix norm subordinate to the Euclidean vector norm on Rm . Indeed, when (7) holds,
the Mean Value Theorem implies that (6) is also valid. The converse of this statement is
not true; for the function f (y) = (|y1 |, . . . , |ym |)T , with x0 = 0 and y0 = 0, satisfies (6)
but violates (7) because y 7→ f (y) is not differentiable at the point y = 0.
As the counter-example in the footnote on page 1 indicates, the expression |y − z|
in (3) and ky − zk in (6) cannot be replaced by expressions of the form |y − z|α and
ky − zkα , respectively, where 0 < α < 1, for otherwise the uniqueness of the solution to
the corresponding initial value problem cannot be guaranteed.
We conclude this section by introducing the notion of stability.
Definition 1 A solution y = v(x) to (5) is said to be stable on the interval [x0 , XM ]
if for every ǫ > 0 there exists δ > 0 such that for all z satisfying kv(x0 ) − zk < δ the
solution y = w(x) to the differential equation y′ = f (x, y) satisfying the initial condition
w(x0 ) = z is defined for all x ∈ [x0 , XM ] and satisfies kv(x) − w(x)k < ǫ for all x in
[x0 , XM ].
A solution which is stable on [x0 , ∞) (i.e. stable on [x0 , XM ] for each XM and with δ
independent of XM ) is said to be stable in the sense of Lyapunov.
Moreover, if
lim kv(x) − w(x)k = 0 ,
x→∞
then the solution y = v(x) is called asymptotically stable.
Using this definition, we can state the following theorem.
Theorem 3 Under the hypotheses of Picard’s theorem, the (unique) solution y = v(x)
to the initial value problem (5) is stable on the interval [x0 , XM ], (where we assume that
−∞ < x0 < XM < ∞).
Proof: Since
v(x) = v(x0 ) +
and
w(x) = z +
Z
Z
x
f (ξ, v(ξ)) dξ
x0
x
f (ξ, w(ξ)) dξ ,
x0
it follows that
kv(x) − w(x)k ≤ kv(x0 ) − zk +
Z
x
kf (ξ, v(ξ)) − f (ξ, w(ξ))k dξ
x0
≤ kv(x0 ) − zk + L
Z
x
x0
kv(ξ) − w(ξ)k dξ .
(8)
Now put A(x) = kv(x) − w(x)k and a = kv(x0 ) − zk; then, (8) can be written as
A(x) ≤ a + L
Z
x
A(ξ) dξ ,
x0 ≤ x ≤ XM .
x0
(9)
Multiplying (9) by exp(−Lx), we find that
d −Lx
e
dx
Z
x
x0
A(ξ) dξ ≤ ae−Lx .
3
(10)
Integrating the inequality (10), we deduce that
−Lx
e
x
Z
x0
that is
L
Z
x
x0
A(ξ) dξ ≤
a −Lx0
e
− e−Lx ,
L
A(ξ) dξ ≤ a eL(x−x0 ) − 1 .
Now substituting (11) into (9) gives
A(x) ≤ aeL(x−x0 ) ,
x0 ≤ x ≤ XM .
(11)
(12)
The implication “(9) ⇒ (12)” is usually referred to as the Gronwall Lemma. Returning
to our original notation, we conclude from (12) that
kv(x) − w(x)k ≤ kv(x0 ) − zkeL(x−x0 ) ,
x0 ≤ x ≤ XM .
(13)
Thus, given ǫ > 0 as in Definition 1, we choose δ = ǫ exp(−L(XM −x0 )) to deduce stability.
⋄
To conclude this section, we observe that if either x0 = −∞ or XM = +∞, the
statement of Theorem 3 is false. For example, the trivial solution y ≡ 0 to the differential
equation y ′ = y is unstable on [x0 , ∞) for any x0 > −∞. More generally, given the initial
value problem
y ′ = λy ,
y(x0 ) = y0 ,
with λ a complex number, the solution y(x) = y0 exp(λ(x − x0 )) is unstable for ℜλ > 0;
the solution is stable in the sense of Lyapunov for ℜλ ≤ 0 and is asymptotically stable for
ℜλ < 0.
In the next section we shall consider numerical methods for the approximate solution
of the initial value problem (1–2). Since everything we shall say has a straightforward
extension to the case of the system (5), for the sake of notational simplicity we shall restrict
ourselves to considering a single ordinary differential equation corresponding to m = 1. We
shall suppose throughout that the function f satisfies the conditions of Picard’s Theorem
on the rectangle R and that the initial value problem has a unique solution defined on the
interval [x0 , XM ], −∞ < x0 < XM < ∞. We begin by discussing one-step methods; this
will be followed in subsequent sections by the study of linear multi-step methods.
2
One-step methods
2.1
Euler’s method and its relatives: the θ-method
The simplest example of a one-step method for the numerical solution of the initial value
problem (1–2) is Euler’s method4 .
Euler’s method. Suppose that the initial value problem (1–2) is to be solved on the
interval [x0 , XM ]. We divide this interval by the mesh-points xn = x0 +nh, n = 0, . . . , N ,
where h = (XM − x0 )/N and N is a positive integer. The positive real number h is called
the step size. Now let us suppose that, for each n, we seek a numerical approximation yn
to y(xn ), the value of the analytical solution at the mesh point xn . Given that y(x0 ) = y0
4
Leonard Euler (1707–1783)
4
is known, let us suppose that we have already calculated yn , up to some n, 0 ≤ n ≤ N − 1;
we define
yn+1 = yn + hf (xn , yn ) ,
n = 0, . . . , N − 1 .
Thus, taking in succession n = 0, 1, . . . , N − 1, one step at a time, the approximate values
yn at the mesh points xn can be easily obtained. This numerical method is known as
Euler’s method.
A simple derivation of Euler’s method proceeds by first integrating the differential
equation (1) between two consecutive mesh points xn and xn+1 to deduce that
y(xn+1 ) = y(xn ) +
Z
xn+1
f (x, y(x)) dx ,
xn
n = 0, . . . , N − 1 ,
(14)
and then applying the numerical integration rule
Z
xn+1
xn
g(x) dx ≈ hg(xn ) ,
called the rectangle rule, with g(x) = f (x, y(x)), to get
y(xn+1 ) ≈ y(xn ) + hf (xn , y(xn )) ,
n = 0, . . . N − 1 ,
y(x0 ) = y0 .
This then motivates the definition of Euler’s method. The idea can be generalised by
replacing the rectangle rule in the derivation of Euler’s method with a one-parameter
family of integration rules of the form
Z
xn+1
xn
g(x) dx ≈ h [(1 − θ)g(xn ) + θg(xn+1 )] ,
(15)
with θ ∈ [0, 1] a parameter. On applying this in (14) with g(x) = f (x, y(x)) we find that
y(xn+1 ) ≈ y(xn ) + h [(1 − θ)f (xn , y(xn )) + θf (xn+1 , y(xn+1 ))] , n = 0, . . . , N − 1 ,
y(x0 ) = y0 .
This then motivates the introduction of the following one-parameter family of methods:
given that y0 is supplied by (2), define
yn+1 = yn + h [(1 − θ)f (xn , yn ) + θf (xn+1 , yn+1 )] ,
n = 0, . . . , N − 1 .
(16)
parametrised by θ ∈ [0, 1]; (16) is called the θ-method. Now, for θ = 0 we recover Euler’s
method. For θ = 1, and y0 specified by (2), we get
yn+1 = yn + hf (xn+1 , yn+1 ) ,
n = 0, . . . , N − 1 ,
(17)
referred to as the Implicit Euler Method since, unlike Euler’s method considered above,
(17) requires the solution of an implicit equation in order to determine yn+1 , given yn .
In order to emphasize this difference, Euler’s method is sometimes termed the Explicit
Euler Method. The scheme which results for the value of θ = 1/2 is also of interest: y0
is supplied by (2) and subsequent values yn+1 are computed from
1
yn+1 = yn + h [f (xn , yn ) + f (xn+1 , yn+1 )] ,
2
5
n = 0, . . . , N − 1 ;
k
0
1
2
3
4
xk
0
0.1
0.2
0.3
0.4
yk for θ = 0
0
0
0.01000
0.02999
0.05990
yk for θ = 1/2
0
0.00500
0.01998
0.04486
0.07944
yk for θ = 1
0
0.00999
0.02990
0.05955
0.09857
Table 1: The values of the numerical solution at the mesh points
this is called the Trapezium Rule Method.
The θ-method is an explicit method for θ = 0 and is an implicit method for 0 < θ ≤ 1,
given that in the latter case (16) requires the solution of an implicit equation for yn+1 .
Further, for each value of the parameter θ ∈ [0, 1], (16) is a one-step method in the sense
that to compute yn+1 we only use one previous value yn . Methods which require more
than one previously computed value are referred to as multi-step methods, and will be
discussed later on in the notes.
In order to assess the accuracy of the θ-method for various values of the parameter θ
in [0, 1], we perform a numerical experiment on a simple model problem.
Example 1 Given the initial value problem y ′ = x − y 2 , y(0) = 0, on the interval of
x ∈ [0, 0.4], we compute an approximate solution using the θ-method, for θ = 0, θ = 1/2
and θ = 1, using the step size h = 0.1. The results are shown in Table 1. In the case
of the two implicit methods, corresponding to θ = 1/2 and θ = 1, the nonlinear equations
have been solved by a fixed-point iteration.
For comparison, we also compute the value of the analytical solution y(x) at the mesh
points xn = 0.1 ∗ n, n = 0, . . . , 4. Since the solution is not available in closed form,5 we
use a Picard iteration to calculate an accurate approximation to the analytical solution on
the interval [0, 0.4] and call this the “exact solution”. Thus, we consider
y0 (x) ≡ 0 ,
yk (x) =
Hence,
Z
0
x
2
ξ − yk−1
(ξ) dξ ,
k = 1, 2, . . . .
y0 (x) ≡ 0 ,
1 2
x ,
y1 (x) =
2
1 2
1
y2 (x) =
x − x5 ,
2
20
5
>
Using MAPLE V, we obtain the solution in terms of Bessel functions:
dsolve({diff(y(x),x)+ y(x)*y(x) = x, y(0)=0}, y(x));
y(x) = −
√
−2 2
3 BesselK( , x3/2 )

√ 
−2 2
3 3
x
− BesselI( , x3/2 )


π
3 3
√
1 2
3 BesselK( , x3/2 )
1 2
3 3
+ BesselI( , x3/2 )
π
3 3
6

k
0
1
2
3
4
xk
0
0.1
0.2
0.3
0.4
y(xk )
0
0.00500
0.01998
0.04488
0.07949
Table 2: Values of the “exact solution” at the mesh points
y3 (x) =
1 2
1
1 8
1 11
x − x5 +
x −
x .
2
20
160
4400
It is easy to prove by induction that
1
1
1 8
1 11
y(x) = x2 − x5 +
x −
x + O x14 ,
2
20
160
4400
Tabulating y3 (x) on the interval [0, 0.4] with step size h = 0.1, we get the “exact solution”
at the mesh points shown in Table 2.
The “exact solution” is in good agreement with the results obtained with θ = 1/2: the
error is ≤ 5 ∗ 10−5 . For θ = 0 and θ = 1 the discrepancy between yk and y(xk ) is larger: it
is ≤ 3 ∗ 10−2 . We note in conclusion that a plot of the analytical solution can be obtained,
for example, by using the MAPLE V package by typing in the following at the command
line:
> with(DEtools):
> DEplot(diff(y(x),x)+y(x)*y(x)=x, y(x), x=0..0.4, [[y(0)=0]],
y=-0.1..0.1, stepsize=0.05);
So, why is the gap between the analytical solution and its numerical approximation in
this example so much larger for θ 6= 1/2 than for θ = 1/2? The answer to this question is
the subject of the next section.
2.2
Error analysis of the θ-method
First we have to explain what we mean by error. The exact solution of the initial value
problem (1–2) is a function of a continuously varying argument x ∈ [x0 , XM ], while the
numerical solution yn is only defined at the mesh points xn , n = 0, . . . , N , so it is a function
of a “discrete” argument. We can compare these two functions either by extending in some
fashion the approximate solution from the mesh points to the whole of the interval [x0 , XM ]
(say by interpolating between the values yn ), or by restricting the function y to the mesh
points and comparing y(xn ) with yn for n = 0, . . . , N . Since the first of these approaches
is somewhat arbitrary because it does not correspond to any procedure performed in a
practical computation, we adopt the second approach, and we define the global error e
by
en = y(xn ) − yn ,
n = 0, . . . , N .
We wish to investigate the decay of the global error for the θ-method with respect to the
reduction of the mesh size h. We shall show in detail how this is done in the case of Euler’s
7
method (θ = 0) and then quote the corresponding result in the general case (0 ≤ θ ≤ 1)
leaving it to the reader to fill the gap.
So let us consider Euler’s explicit method:
yn+1 = yn + hf (xn , yn ) ,
n = 0, . . . , N − 1 ,
y0 = given .
The quantity
y(xn+1 ) − y(xn )
(18)
− f (xn , y(xn )) ,
h
obtained by inserting the analytical solution y(x) into the numerical method and dividing
by the mesh size is referred to as the truncation error of Euler’s explicit method and
will play a key role in the analysis. Indeed, it measures the extent to which the analytical
solution fails to satisfy the difference equation for Euler’s method.
By noting that f (xn , y(xn )) = y ′ (xn ) and applying Taylor’s Theorem, it follows from
(18) that there exists ξn ∈ (xn , xn+1 ) such that
Tn =
1 ′′
hy (ξn ) ,
(19)
2
where we have assumed that that f is a sufficiently smooth function of two variables so as
to ensure that y ′′ exists and is bounded on the interval [x0 , XM ]. Since from the definition
of Euler’s method
yn+1 − yn
0=
− f (xn , yn ) ,
h
on subtracting this from (18), we deduce that
Tn =
en+1 = en + h[f (xn , y(xn )) − f (xn , yn )] + hTn .
Thus, assuming that |yn − y0 | ≤ YM from the Lipschitz condition (3) we get
|en+1 | ≤ (1 + hL)|en | + h|Tn | ,
n = 0, . . . , N − 1 .
Now, let T = max0≤n≤N −1 |Tn | ; then,
|en+1 | ≤ (1 + hL)|en | + hT ,
n = 0, . . . , N − 1 .
By induction, and noting that 1 + hL ≤ ehL ,
|en | ≤
≤
T
[(1 + hL)n − 1] + (1 + hL)n |e0 |
L
T L(xn −x0 )
e
− 1 + eL(xn −x0 ) |e0 | ,
L
n = 1, . . . , N .
This estimate, together with the bound
|T | ≤
1
hM2 ,
2
M2 =
max
x∈[x0 ,XM ]
|y ′′ (x)| ,
which follows from (19), yields
|en | ≤ eL(xn −x0 ) |e0 | +
M2 h L(xn −x0 )
e
−1 ,
2L
8
n = 0, . . . , N .
(20)
To conclude, we note that pursuing an analogous argument it is possible to prove that,
in the general case of the θ-method,
xn − x0
|en | ≤ |e0 | exp L
1 − θLh
h 1
1
xn − x0
− θ M2 + hM3 exp L
+
−1 ,
L 2
3
1 − θLh
(21)
for n = 0, . . . , N , where now M3 = maxx∈[x0 ,XM ] |y ′′′ (x)|. In the absence of rounding errors
in the imposition of the initial condition (2) we can suppose that e0 = y(x0 ) − y0 = 0.
Assuming that this is the case, we see from (21) that |en | = O(h2 ) for θ = 1/2, while for
θ = 0 and θ = 1, and indeed for any θ 6= 1/2, |en | = O(h) only. This explains why in
Tables 1 and 2 the values yn of the numerical solution computed with the trapezium-rule
method (θ = 1/2) were considerably closer to the analytical solution y(xn ) at the mesh
points than those which were obtained with the explicit and the implicit Euler methods
(θ = 0 and θ = 1, respectively).
In particular, we see from this analysis, that each time the mesh size h is halved, the
truncation error and the global error are reduced by a factor of 2 when θ 6= 1/2, and by a
factor of 4 when θ = 1/2.
While the trapezium rule method leads to more accurate approximations than the
forward Euler method, it is less convenient from the computational point of view given
that it requires the solution of implicit equations at each mesh point xn+1 to compute
yn+1 . An attractive compromise is to use the forward Euler method to compute an initial
crude approximation to y(xn+1 ) and then use this value within the trapezium rule to
obtain a more accurate approximation for y(xn+1 ): the resulting numerical method is
1
yn+1 = yn + h [f (xn , yn ) + f (xn+1 , yn + hf (xn , yn ))] ,
2
n = 0, . . . , N ,
y0 = given ,
and is frequently referred to as the improved Euler method. Clearly, it is an explicit
one-step scheme, albeit of a more complicated form than the explicit Euler method. In the
next section, we shall take this idea further and consider a very general class of explicit
one-step methods.
2.3
General explicit one-step method
A general explicit one-step method may be written in the form:
yn+1 = yn + hΦ(xn , yn ; h) ,
n = 0, . . . , N − 1 ,
y0 = y(x0 ) [= specified by (2)] , (22)
where Φ(·, ·; ·) is a continuous function of its variables. For example, in the case of Euler’s
method, Φ(xn , yn ; h) = f (xn , yn ), while for the improved Euler method
Φ(xn , yn ; h) =
1
[f (xn , yn ) + f (xn + h, yn + hf (xn , yn ))] .
2
In order to assess the accuracy of the numerical method (22), we define the global
error, en , by
en = y(xn ) − yn .
9
We define the truncation error, Tn , of the method by
Tn =
y(xn+1 ) − y(xn )
− Φ(xn , y(xn ); h) .
h
(23)
The next theorem provides a bound on the global error in terms of the truncation
error.
Theorem 4 Consider the general one-step method (22) where, in addition to being a
continuous function of its arguments, Φ is assumed to satisfy a Lipschitz condition with
respect to its second argument; namely, there exists a positive constant LΦ such that, for
0 ≤ h ≤ h0 and for the same region R as in Picard’s theorem,
|Φ(x, y; h) − Φ(x, z; h)| ≤ LΦ |y − z|,
for (x, y), (x, z) in R .
(24)
#
(25)
Then, assuming that |yn − y0 | ≤ YM , it follows that
|en | ≤ e
LΦ (xn −x0 )
"
eLΦ (xn −x0 ) − 1
|e0 | +
T ,
LΦ
n = 0, . . . , N ,
where T = max0≤n≤N −1 |Tn | .
Proof: Subtracting (22) from (23) we obtain:
en+1 = en + h[Φ(xn , y(xn ); h) − Φ(xn , yn ; h)] + hTn .
Then, since (xn , y(xn )) and (xn , yn ) belong to R, the Lipschitz condition (24) implies that
|en+1 | ≤ |en | + hLΦ |en | + h|Tn | ,
n = 0, . . . , N − 1 .
|en+1 | ≤ (1 + hLΦ )|en | + h|Tn | ,
n = 0, . . . , N − 1 .
That is,
Hence
|e1 |
≤
(1 + hLΦ )|e0 | + hT ,
|e3 |
≤
(1 + hLΦ )3 |e0 | + h[1 + (1 + hLΦ ) + (1 + hLΦ )2 ]T ,
|e2 |
|en |
≤
etc.
≤
(1 + hLΦ )2 |e0 | + h[1 + (1 + hLΦ )]T ,
(1 + hLΦ )n |e0 | + [(1 + hLΦ )n − 1]T /LΦ .
Observing that 1 + hLΦ ≤ exp(hLΦ ), we obtain (25). ⋄
Let us note that the error bound (20) for Euler’s explicit method is a special case
of (25). We highlight the practical relevance of the error bound (25) by focusing on a
particular example.
Example 2 Consider the initial value problem y ′ = tan−1 y, y(0) = y0 , and suppose that
this is solved by the explicit Euler method. The aim of the exercise is to apply (25) to
10
quantify the size of the associated global error; thus, we need to find L and M2 . Here
f (x, y) = tan−1 y, so by the Mean Value Theorem
∂f
|f (x, y) − f (x, z)| = (x, η) (y − z) ,
∂y
where η lies between y and z. In our case
∂f 2 −1
∂y = |(1 + y ) | ≤ 1 ,
and therefore L = 1. To find M2 we need to obtain a bound on |y ′′ | (without actually
solving the initial value problem!). This is easily achieved by differentiating both sides of
the differential equation with respect to x:
y ′′ =
d
dy
(tan−1 y) = (1 + y 2 )−1
= (1 + y 2 )−1 tan−1 y .
dx
dx
Therefore |y ′′ (x)| ≤ M2 = 12 π. Inserting the values of L and M2 into (20),
1
|en | ≤ exn |e0 | + π (exn − 1) h ,
4
n = 0, . . . , N .
In particular if we assume that no error has been committed initially (i.e. e0 = 0), we
have that
1
|en | ≤ π (exn − 1) h , n = 0, . . . , N .
4
Thus, given a tolerance T OL specified beforehand, we can ensure that the error between
the (unknown) analytical solution and its numerical approximation does not exceed this
tolerance by choosing a positive step size h such that
h≤
4 XM
(e
− 1)−1 T OL .
π
For such h we shall have |y(xn ) − yn | = |en | ≤ T OL for each n = 0, . . . , N , as required.
Thus, at least in principle, we can calculate the numerical solution to arbitrarily high accuracy by choosing a sufficiently small step size. In practice, because digital computers use
finite-precision arithmetic, there will always be small (but not infinitely small) pollution effects due to rounding errors; however, these can also be bounded by performing an analysis
similar to the one above where f (xn , yn ) is replaced by its finite-precision representation.
Returning to the general one-step method (22), we consider the choice of the function
Φ. Theorem 4 suggests that if the truncation error ‘approaches zero’ as h → 0 then the
global error ‘converges to zero’ also (as long as |e0 | → 0 when h → 0). This observation
motivates the following definition.
Definition 2 The numerical method (22) is consistent with the differential equation
(1) if the truncation error defined by (23) is such that for any ǫ > 0 there exists a positive
h(ǫ) for which |Tn | < ǫ for 0 < h < h(ǫ) and any pair of points (xn , y(xn )), (xn+1 , y(xn+1 ))
on any solution curve in R.
11
For the general one-step method (22) we have assumed that the function Φ(·, ·; ·) is
continuous; also y ′ is a continuous function on [x0 , XM ]. Therefore, from (23),
lim Tn = y ′ (xn ) − Φ(xn , y(xn ); 0) .
h→0
This implies that the one-step method (22) is consistent if and only if
Φ(x, y; 0) ≡ f (x, y) .
(26)
Now we are ready to state a convergence theorem for the general one-step method (22).
Theorem 5 Suppose that the solution of the initial value problem (1–2) lies in R as does
its approximation generated from (22) when h ≤ h0 . Suppose also that the function Φ(·, ·; ·)
is uniformly continuous on R × [0, h0 ] and satisfies the consistency condition (26) and the
Lipschitz condition
|Φ(x, y; h) − Φ(x, z; h)| ≤ LΦ |y − z|
on R × [0, h0 ] .
(27)
Then, if successive approximation sequences (yn ), generated for xn = x0 + nh, n =
1, 2, . . . , N , are obtained from (22) with successively smaller values of h, each less than h0 ,
we have convergence of the numerical solution to the solution of the initial value problem
in the sense that
|y(xn ) − yn | → 0
as h → 0, xn → x ∈ [x0 , XM ] .
Proof: Suppose that h = (XM − x0 )/N where N is a positive integer. We shall assume
that N is sufficiently large so that h ≤ h0 . Since y(x0 ) = y0 and therefore e0 = 0, Theorem
4 implies that
"
eLφ (XM −x0 ) − 1
|y(xn ) − yn | ≤
Lφ
#
max
0≤m≤n−1
|Tm | ,
n = 1, . . . , N .
(28)
From the consistency condition (26) we have
y(xn+1 ) − y(xn )
Tn =
− f (xn , y(xn )) + [Φ(xn , y(xn ); 0) − Φ(xn , y(xn ); h)] .
h
According to the Mean Value Theorem the expression in the first bracket is equal to
y ′ (ξ) − y ′ (xn ), where ξ ∈ [xn , xn+1 ]. Since y ′ (·) = f (·, y(·)) = Φ(·, y(·); 0) and Φ(·, ·; ·) is
uniformly continuous on R × [0, h0 ], it follows that y ′ is uniformly continuous on [x0 , XM ].
Thus, for each ǫ > 0 there exists h1 (ǫ) such that
1
|y ′ (ξ) − y ′ (xn )| ≤ ǫ
2
for h < h1 (ǫ), n = 0, 1, . . . , N − 1 .
Also, by the uniform continuity of Φ with respect to its third argument, there exists h2 (ǫ)
such that
|Φ(xn , y(xn ); 0) − Φ(xn , y(xn ); h)| ≤
1
ǫ
2
12
for h < h2 (ǫ), n = 0, 1, . . . , N − 1 .
Thus, defining h(ǫ) = min(h1 (ǫ), h2 (ǫ)), we have
|Tn | ≤ ǫ
for h < h(ǫ), n = 0, 1, . . . , N − 1 .
Inserting this into (28) we deduce that |y(xn ) − yn | → 0 as h → 0. Since
|y(x) − yn | ≤ |y(x) − y(xn )| + |y(xn ) − yn | ,
and the first term on the right also converges to zero as h → 0 by the uniform continuity
of y on the interval [x0 , XM ], the proof is complete. ⋄
We saw earlier that for Euler’s method the absolute value of the truncation error Tn
is bounded above by a constant multiple of the step size h, that is
|Tn | ≤ Kh
for 0 < h ≤ h0 ,
where K is a positive constant, independent of h. However there are other one-step
methods (a class of which, called Runge–Kutta methods, will be considered below) for
which we can do better. More generally, in order to quantify the asymptotic rate of decay
of the truncation error as the step size h converges to zero, we introduce the following
definition.
Definition 3 The numerical method (22) is said to have order of accuracy p, if p is
the largest positive integer such that, for any sufficiently smooth solution curve (x, y(x))
in R of the initial value problem (1–2), there exist constants K and h0 such that
|Tn | ≤ Khp
for 0 < h ≤ h0
for any pair of points (xn , y(xn )), (xn+1 , y(xn+1 )) on the solution curve.
Having introduced the general class of explicit one-step methods and the associated
concepts of consistency and order of accuracy, we now focus on a specific family: explicit
Runge–Kutta methods.
2.4
Runge–Kutta methods
In the sense of Definition 3 Euler’s method is only first-order accurate; nevertheless, it
is simple and cheap to implement because to obtain yn+1 from yn we only require a
single evaluation of the function f at (xn , yn ). Runge–Kutta methods aim to achieve
higher accuracy by sacrificing the efficiency of Euler’s method through re-evaluating f (·, ·)
at points intermediate between (xn , y(xn )) and (xn+1 , y(xn+1 )). The general R-stage
Runge–Kutta family is defined by
yn+1 = yn + hΦ(xn , yn ; h) ,
Φ(x, y; h) =
R
X
cr kr ,
r=1
k1 = f (x, y) ,
kr = f x + har , y + h
r−1
X
brs ks
s=1
ar =
r−1
X
brs ,
r = 2, . . . , R .
s=1
13
!
,
r = 2, . . . , R ,
(29)
a = Be
B
where e = (1, . . . , 1)T .
cT
Figure 1: Butcher table of a Runge–Kutta method
In compressed form, this information is usually displayed in the so-called Butcher table
displayed in Figure 1.
One-stage Runge–Kutta methods. Suppose that R = 1. Then, the resulting onestage Runge–Kutta method is simply Euler’s explicit method:
yn+1 = yn + hf (xn , yn ) .
(30)
Two-stage Runge–Kutta methods. Next, consider the case of R = 2, corresponding
to the following family of methods:
yn+1 = yn + h(c1 k1 + c2 k2 ) ,
(31)
where
k1 = f (xn , yn ) ,
(32)
k2 = f (xn + a2 h, yn + b21 hk1 ) ,
(33)
and where the parameters c1 , c2 , a2 and b21 are to be determined.6 Clearly (31–33) can
be rewritten in the form (22) and therefore it is a family of one step methods. By the
condition (26), a method from this family will be consistent if and only if
c1 + c2 = 1 .
Further conditions on the parameters are obtained by attempting to maximise the order
of accuracy of the method. Indeed, expanding the truncation error of (31–33) in powers
of h, after some algebra we obtain
Tn =
1 ′′
1
hy (xn ) + h2 y ′′′ (xn )
2
6
−c2 h[a2 fx + b21 fy f ] − c2 h2
1 2
1
a2 fxx + a2 b21 fxy f + b221 fyy f 2 + O(h3 ) .
2
2
Here we have used the abbreviations f = f (xn , y(xn )), fx = ∂f
∂x (xn , y(xn )), etc. On noting
that y ′′ = fx + fy f , it follows that Tn = O(h2 ) for any f provided that
a2 c2 = b21 c2 =
1
,
2
which implies that if b21 = a2 , c2 = 1/(2a2 ) and c1 = 1 − 1/(2a2 ) then the method is
second-order accurate; while this still leaves one free parameter, a2 , it is easy to see that
no choice of the parameters will make the method generally third-order accurate. There
are two well-known examples of second-order Runge–Kutta methods of the form (31–33):
6
We note in passing that Euler’s method is a member of this family of methods, corresponding to c1 = 1
and c2 = 0. However we are now seeking methods that are at least second-order accurate.
14
a) The modified Euler method: In this case we take a2 =
1
2
to obtain
1
1
yn+1 = yn + h f xn + h, yn + hf (xn , yn )
2
2
;
b) The improved Euler method: This is arrived at by choosing a2 = 1 which gives
1
yn+1 = yn + h [f (xn , yn ) + f (xn + h, yn + hf (xn , yn ))] .
2
For these two methods it is easily verified by Taylor series expansion that the truncation
error is of the form, respectively,
Tn =
Tn =
1 2
h fy F1 +
6
1 2
h fy F1 −
6
1
F2 + O(h3 ) ,
4
1
F2 + O(h3 ) ,
2
where
F1 = fx + f fy
and F2 = fxx + 2f fxy + f 2 fyy .
The family (31–33) is referred to as the class of explicit two-stage Runge–Kutta methods.
Exercise 1 Let α be a non-zero real number and let xn = a + nh, n = 0, . . . , N , be a
uniform mesh on the interval [a, b] of step size h = (b − a)/N . Consider the explicit onestep method for the numerical solution of the initial value problem y ′ = f (x, y), y(a) = y0 ,
which determines approximations yn to the values y(xn ) from the recurrence relation
yn+1 = yn + h(1 − α)f (xn , yn ) + hαf xn +
h
h
, yn +
f (xn , yn ) .
2α
2α
Show that this method is consistent and that its truncation error, Tn (h, α), can be
expressed as
Tn (h, α) =
h2
8α
∂f
4
α − 1 y ′′′ (xn ) + y ′′ (xn ) (xn , y(xn )) + O(h3 ) .
3
∂y
This numerical method is applied to the initial value problem y ′ = −y p , y(0) = 1, where
p is a positive integer. Show that if p = 1 then Tn (h, α) = O(h2 ) for every non-zero real
number α. Show also that if p ≥ 2 then there exists a non-zero real number α0 such that
Tn (h, α0 ) = O(h3 ).
Solution: Let us define
Φ(x, y; h) = (1 − α)f (x, y) + αf
x+
h
h
,y +
f (x, y) .
2α
2α
Then the numerical method can be rewritten as
yn+1 = yn + hΦ(xn , yn ; h) .
Since
Φ(x, y; 0) = f (x, y) ,
15
the method is consistent. By definition, the truncation error is
Tn (h, α) =
y(xn+1 ) − y(xn )
− Φ(xn , y(xn ); h) .
h
We shall perform a Taylor expansion of Tn (h, α) to show that it can be expressed in the desired
form. Indeed,
Tn (h, α)
h2
h ′′
y (xn ) + y ′′′ (xn )
2
6
h
h ′
′
−(1 − α)y (xn ) − αf (xn +
, y(xn ) +
y (xn )) + O(h3 )
2α
2α
h
h2
= y ′ (xn ) + y ′′ (xn ) + y ′′′ (xn ) − (1 − α)y ′ (xn )
2
6
h
h
′
−α f (xn , y(xn )) +
fx (xn , y(xn )) +
fy (xn , y(xn ))y (xn )
2α
2α
" 2
2
α
h
h
−
fxx (xn , y(xn )) + 2
fxy (xn , y(xn ))y ′ (xn )
2
2α
2α
#
2
h
′
2
fyy (xn , y(xn ))[y (xn )] + O(h3 )
+
2α
= y ′ (xn ) +
= y ′ (xn ) − (1 − α)y ′ (xn ) − αy ′ (xn )
h
h
+ y ′′ (xn ) − [fx (xn , y(xn )) + fy (xn , y(xn ))y ′ (xn )]
2
2
h2 ′′′
h2
+ y (xn ) −
[fxx (xn , y(xn )) + 2fxy (xn , y(xn ))y ′ (xn )
6
8α
+ fyy (xn , y(xn ))[y ′ (xn )]2 + O(h3 )
=
=
h2 ′′′
h2 ′′′
y (xn ) −
[y (xn ) − y ′′ (xn )fy (xn , y(xn ))] + O(h3 )
6 8α
h2
4
∂f
α − 1 y ′′′ (xn ) + y ′′ (xn ) (xn , y(xn )) + O(h3 ) ,
8α
3
∂y
as required.
Now let us apply the method to y ′ = −y p , with p ≥ 1. If p = 1, then y ′′′ = −y ′′ = y ′ = −y, so
that
h2
Tn (h, α) = − y(xn ) + O(h3 ) .
6
−xn
As y(xn ) = e
6= 0, it follows that
Tn (h, α) = O(h2 )
for all (non-zero) α.
Finally, suppose that p ≥ 2. Then
y ′′ = −py p−1 y ′ = py 2p−1
and
y ′′′ = p(2p − 1)y 2p−2 y ′ = −p(2p − 1)y 3p−2 ,
and therefore
Tn (h, α) = −
h2
8α
4
α − 1 p(2p − 1) + p2 y 3p−2 (xn ) + O(h3 ) .
3
16
Choosing α such that
4
α − 1 p(2p − 1) + p2 = 0 ,
3
namely
α = α0 =
gives
3p − 3
,
8p − 4
Tn (h, α0 ) = O(h3 ) .
We note in passing that for p > 1 the exact solution of the initial value problem y ′ = −y p ,
1/(1−p)
y(0) = 1, is y(x) = [(p − 1)x + 1]
. ⋄
Three-stage Runge–Kutta methods. Let us now suppose that R = 3 to illustrate
the general idea. Thus, we consider the family of methods:
yn+1 = yn + h [c1 k1 + c2 k2 + c3 k3 ] ,
where
k1 = f (x, y) ,
k2 = f (x + ha2 , y + hb21 k1 ) ,
k3 = f (x + ha3 , y + hb31 k1 + hb32 k2 ) ,
a2 = b21 ,
a3 = b31 + b32 .
Writing b21 = a2 and b31 = a3 − b32 in the definitions of k2 and k3 respectively and
expanding k2 and k3 into Taylor series about the point (x, y) yields:
1
k2 = f + ha2 (fx + k1 fy ) + h2 a22 (fxx + 2k1 fxy + k12 fyy ) + O(h3 )
2
1 2 2
= f + ha2 (fx + f fy ) + h a2 (fxx + 2f fxy + f 2 fyy ) + O(h3 )
2
1 2 2
= f + ha2 F1 + h a2 F2 + O(h3 ) ,
2
where
F1 = fx + f fy
and F2 = fxx + 2f fxy + f 2 fyy ,
and
k3 = f + h {a3 fx + [(a3 − b32 )k1 + b32 k2 ] fy }
1 n
+ h2 a23 fxx + 2a3 [(a3 − b32 )k1 + b32 k2 ] fxy
2
o
+ [(a3 − b32 )k1 + b32 k2 ]2 fyy + O(h3 )
1
= f + ha3 F1 + h2 a2 b32 F1 fy + a23 F2 + O(h3 ) .
2
Substituting these expressions for k2 and k3 into (29) with R = 3 we find that
Φ(x, y, h) = (c1 + c2 + c3 )f + h(c2 a2 + c3 a3 )F1
i
1 h
+ h2 2c3 a2 b32 F1 fy + c2 a22 + c3 a23 F2 + O(h3 ) .
2
17
(34)
We match this with the Taylor series expansion:
y(x + h) − y(x)
h
1
1
= y ′ (x) + hy ′′ (x) + h2 y ′′′ (x) + O(h3 )
2
6
1 2
1
= f + hF1 + h (F1 fy + F2 ) + O(h3 ) .
2
6
This yields:
c1 + c2 + c3 = 1 ,
1
c2 a2 + c3 a3 =
,
2
1
c2 a22 + c3 a23 =
,
3
1
.
c3 a2 b32 =
6
Solving this system of four equations for the six unknowns: c1 , c2 , c3 , a2 , a3 , b32 , we obtain
a two-parameter family of 3-stage Runge–Kutta methods. We shall only highlight two
notable examples from this family:
(i) Heun’s method corresponds to
c1 =
1
,
4
c2 = 0 ,
c3 =
3
,
4
a2 =
1
,
3
a3 =
2
,
3
b32 =
2
,
3
yielding
1
yn+1 = yn + h (k1 + 3k3 ) ,
4
k1 = f (xn , yn ) ,
1
1
k2 = f xn + h, yn + hk1 ,
3
3
2
2
k3 = f xn + h, yn + hk2 .
3
3
(ii) Standard third-order Runge–Kutta method. This is arrived at by selecting
c1 =
1
,
6
c2 =
2
,
3
c3 =
1
,
6
a2 =
1
,
2
a3 = 1 ,
yielding
1
yn+1 = yn + h (k1 + 4k2 + k3 ) ,
6
k1 = f (xn , yn ) ,
1
1
k2 = f xn + h, yn + hk1 ,
2
2
k3 = f (xn + h, yn − hk1 + 2hk2 ) .
18
b32 = 2 ,
Four-stage Runge–Kutta methods. For R = 4, an analogous argument leads to a
two-parameter family of four-stage Runge–Kutta methods of order four. A particularly
popular example from this family is:
1
yn+1 = yn + h (k1 + 2k2 + 2k3 + k4 ) ,
6
where
k1 = f (xn , yn ) ,
1
1
k2 = f xn + h, yn + hk1 ,
2
2
1
1
k3 = f xn + h, yn + hk2 ,
2
2
k4 = f (xn + h, yn + hk3 ) .
Here k2 and k3 represent approximations to the derivative y ′ (·) at points on the solution
curve, intermediate between (xn , y(xn )) and (xn+1 , y(xn+1 )), and Φ(xn , yn ; h) is a weighted
average of the ki , i = 1, . . . , 4, the weights corresponding to those of Simpson’s rule method
(to which the fourth-order Runge–Kutta method reduces when ∂f
∂y ≡ 0).
In this section, we have constructed R-stage Runge–Kutta methods of order of accuracy
O(hR ), R = 1, 2, 3, 4. Is is natural to ask whether there exists an R stage method of order
R for R ≥ 5. The answer to this question is negative: in a series of papers John Butcher
showed that for R = 5, 6, 7, 8, 9, the highest order that can be attained by an R-stage
Runge–Kutta method is, respectively, 4, 5, 6, 6, 7, and that for R ≥ 10 the highest order is
≤ R − 2.
2.5
Absolute stability of Runge–Kutta methods
It is instructive to consider the model problem
y ′ = λy ,
y(0) = y0 (6= 0),
(35)
with λ real and negative. Trivially, the analytical solution to this initial value problem,
y(x) = y0 exp(λx), converges to 0 at an exponential rate as x → +∞. The question that
we wish to investigate here is under what conditions on the step size h does a Runge–Kutta
method reproduce this behaviour. The understanding of this matter will provide useful
information about the adequate selection of h in the numerical approximation of an initial
value problem by a Runge–Kutta method over an interval [x0 , XM ] with XM >> x0 . For
the sake of simplicity, we shall restrict our attention to the case of R-stage methods of
order of accuracy R, with 1 ≤ R ≤ 4.
Let us begin with R = 1. The only explicit one-stage first-order accurate Runge–Kutta
method is Euler’s explicit method. Applying (30–35) yields:
¯ n,
yn+1 = (1 + h)y
¯ = λh. Thus,
where h
n≥0,
¯ n y0 .
yn = (1 + h)
19
¯
Consequently, the sequence {yn }∞
n=0 will converge to 0 if and only if |1 + h| < 1, yielding
¯ ∈ (−2, 0); for such h the Euler’s explicit method is said to be absolutely stable and
h
the interval (−2, 0) is referred to as the interval of absolute stability of the method.
Now consider R = 2 corresponding to two-stage second-order Runge–Kutta methods:
yn+1 = yn + h(c1 k1 + c2 k2 ) ,
where
k1 = f (xn , yn ),
k2 = f (xn + a2 h, yn + b21 hk1 )
with
c1 + c2 = 1 ,
a2 c2 = b21 c2 =
1
.
2
Applying this to (35) yields,
yn+1
and therefore
¯ 2 yn ,
¯ + 1h
= 1+h
2
¯2
¯ + 1h
yn = 1 + h
2
n≥0,
n
y0 .
Hence the method is absolutely stable if and only if
¯ + 1h
¯2 < 1 ,
1 + h
2 ¯ ∈ (−2, 0).
namely when h
In the case of R = 3 an analogous argument shows that
¯ + 1h
¯2 +
yn+1 = 1 + h
2
Demanding that
1 ¯3
h yn .
6
¯ + 1h
¯2 + 1 h
¯3 < 1
1 + h
2
6 ¯ ∈ (−2.51, 0).
then yields the interval of absolute stability: h
When R = 4, we have that
yn+1
1 ¯2
= 1+¯
h+ h
+
2
1 ¯3
1 ¯4
h + h
yn ,
6
24
¯ ∈ (−2.78, 0).
and the associated interval of absolute stability is h
For R ≥ 5 on applying the Runge–Kutta method to the model problem (35) still results
in a recursion of the form
¯ n,
yn+1 = AR (h)y
n≥0,
¯ the expression AR (h)
¯ also
however, unlike the case when R = 1, 2, 3, 4, in addition to h
depends on the coefficients of the Runge–Kutta method; by a convenient choice of the free
parameters the associated interval of absolute stability may be maximised. For further
results in this direction, the reader is referred to the book of J.D. Lambert.
20
3
Linear multi-step methods
While Runge–Kutta methods present an improvement over Euler’s method in terms of
accuracy, this is achieved by investing additional computational effort; in fact, Runge–
Kutta methods require more evaluations of f (·, ·) than would seem necessary. For example,
the fourth-order method involves four function evaluations per step. For comparison, by
considering three consecutive points xn−1 , xn = xn−1 +h, xn+1 = xn−1 +2h, integrating the
differential equation between xn−1 and xn+1 , and applying Simpson’s rule to approximate
the resulting integral yields
y(xn+1 ) = y(xn−1 ) +
Z
xn+1
f (x, y(x)) dx
xn−1
1
≈ y(xn−1 ) + h [f (xn−1 , y(xn−1 )) + 4f (xn , y(xn )) + f (xn+1 , y(xn+1 ))]
3
which leads to the method
1
yn+1 = yn−1 + h [f (xn−1 , yn−1 ) + 4f (xn , yn ) + f (xn+1 , yn+1 )] .
3
(36)
In contrast with the one-step methods considered in the previous section where only a
single value yn was required to compute the next approximation yn+1 , here we need two
preceding values, yn and yn−1 to be able to calculate yn+1 , and therefore (36) is not a
one-step method.
In this section we consider a class of methods of the type (36) for the numerical solution
of the initial value problem (1–2), called linear multi-step methods.
Given a sequence of equally spaced mesh points (xn ) with step size h, we consider the
general linear k-step method
k
X
j=0
αj yn+j = h
k
X
βj f (xn+j , yn+j ) ,
(37)
j=0
where the coefficients α0 , . . . , αk and β0 , . . . , βk are real constants. In order to avoid
degenerate cases, we shall assume that αk 6= 0 and that α0 and β0 are not both equal to
zero. If βk = 0 then yn+k is obtained explicitly from previous values of yj and f (xj , yj ),
and the k-step method is then said to be explicit. On the other hand, if βk 6= 0 then yn+k
appears not only on the left-hand side but also on the right, within f (xn+k , yn+k ); due
to this implicit dependence on yn+k the method is then called implicit. The numerical
method (37) is called linear because it involves only linear combinations of the {yn } and
the {f (xn , yn )}; for the sake of notational simplicity, henceforth we shall write fn instead
of f (xn , yn ).
Example 3 We have already seen an example of a linear 2-step method in (36); here we
present further examples of linear multi-step methods.
a) Euler’s method is a trivial case: it is an explicit linear one-step method. The implicit Euler method
yn+1 = yn + hf (xn+1 , yn+1 )
is an implicit linear one-step method.
21
b) The trapezium method, given by
1
yn+1 = yn + h[fn+1 + fn ]
2
is also an implicit linear one-step method.
c) The four-step Adams7 - Bashforth method
1
h[55fn+3 − 59fn+2 + 37fn+1 − 9fn ]
24
is an example of an explicit linear four-step method; the four-step Adams - Moulton method
1
yn+4 = yn+3 + h[9fn+4 + 19fn+3 − 5fn+2 − 9fn+1 ]
24
is an implicit linear four-step method.
yn+4 = yn+3 +
The construction of general classes of linear multi-step methods, such as the (implicit)
Adams–Bashforth family and the (explicit) Adams–Moulton family will be discussed in the
next section.
3.1
Construction of linear multi-step methods
Let us suppose that un , n = 0, 1, . . ., is a sequence of real numbers. We introduce the shift
operator E, the forward difference operator ∆+ and the backward difference operator ∆−
by
E : un 7→ un+1 ,
∆+ : un 7→ (un+1 − un ) ,
∆− : un 7→ (un − un−1 ) .
Further, we note that E −1 exists and is given by E −1 : un+1 7→ un . Since
∆+ = E − I = E∆− ,
∆− = I − E −1
E = (I − ∆− )−1 ,
and
where I signifies the identity operator, it follows that, for any positive integer k,
∆k+ un
k
= (E − I) un =
k
X
= (I − E
(−1)
j=0
and
∆k− un
k
j
j
−1 k
) un =
k
X
j
(−1)
j=0
!
un+k−j
k
j
!
un−j .
Now suppose that u is a real-valued function defined on R whose derivative exists and
is integrable on [x0 , xn ] for each n ≥ 0, and let un denote u(xn ) where xn = x0 + nh,
n = 0, 1, . . ., are equally spaced points on the real line. Letting D denote d/dx, by
applying a Taylor series expansion we find that
(E s u)n = u(xn + sh) = un + sh(Du)n +
=
7
∞
X
1
k!
k=0
1
(sh)2 (D 2 u)n + · · ·
2!
((shD)k u)n = (eshD u)n ,
J. C. Adams (1819–1892)
22
and hence
E s = eshD .
Thus, formally,
hD = lnE = −ln(I − ∆− ) ,
and therefore, again by Taylor series expansion,
1
1
hu (xn ) = ∆− + ∆2− + ∆3− + · · · un .
2
3
′
Now letting u(x) = y(x) where y is the solution of the initial-value problem (1–2) and
noting that u′ (x) = y ′ (x) = f (x, y(x)), we find that
1
1
hf (xn , y(xn )) = ∆− + ∆2− + ∆3− + · · · y(xn ) .
2
3
On successive truncation of the infinite series on the right, we find that
y(xn ) − y(xn−1 ) ≈ hf (xn , y(xn )) ,
3
1
y(xn ) − 2y(xn−1 ) + y(xn−2 ) ≈ hf (xn , y(xn )) ,
2
2
11
3
1
y(xn ) − 3y(xn−1 ) + y(xn−2 ) − y(xn−3 ) ≈ hf (xn , y(xn )) ,
6
2
3
and so on. These approximate equalities give rise to a class of implicit linear multi-step
methods called backward differentiation formulae, the simplest of which is Euler’s
implicit method.
Similarly,
E −1 (hD) = hDE −1 = (I − ∆− )(−ln(I − ∆− )) = −(I − ∆− )ln(I − ∆− ) ,
and therefore
1
1
hu (xn ) = ∆− − ∆2− − ∆3− + · · · un+1 .
2
6
′
Letting, again, u(x) = y(x) where y is the solution of the initial-value problem (1–2) and
noting that u′ (x) = y ′ (x) = f (x, y(x)), on successive truncation of the infinite series on
the right results in
y(xn+1 ) − y(xn ) ≈ hf (xn , y(xn )) ,
1
1
y(xn+1 ) − y(xn−1 ) ≈ hf (xn , y(xn )) ,
2
2
1
1
1
y(xn+1 ) + y(xn ) − y(xn−1 ) + y(xn−2 ) ≈ hf (xn , y(xn )) ,
3
2
6
and so on. The first of these yields Euler’s explicit method, the second the so-called
explicit mid-point rule, and so on.
Next we derive additional identities which will allow us to construct further classes of
linear multi-step methods. Let us define
D −1 u(xn ) = u(x0 ) +
23
Z
xn
x0
u(ξ) dξ ,
and observe that
(E − I)D
−1
u(xn ) =
Now,
Z
xn+1
u(ξ) dξ .
xn
(E − I)D −1 = ∆+ D −1 = E∆− D −1 = hE∆− (hD)−1
= −hE∆− [ln(I − ∆− )]−1 .
(38)
Furthermore,
(E − I)D −1 = E∆− D −1 = ∆− ED −1 = ∆− (DE −1 )−1 = h∆− (hDE −1 )−1
= −h∆− [(I − ∆− )ln(I − ∆− )]−1 .
(39)
Letting, again, u(x) = y(x) where y is the solution of the initial-value problem (1–2),
noting that u′ (x) = y ′ (x) = f (x, y(x)) and using (38) and (39) we deduce that
y(xn+1 ) − y(xn ) =
=
Z
xn+1
xn
(
y ′ (ξ) dξ = (E − I)D −1 y ′ (xn ) = (E − I)D −1 f (xn , y(xn ))
−hE∆− [ln(I − ∆− )]−1 f (xn , y(xn ))
−h∆− [(I − ∆− )ln(I − ∆− )]−1 f (xn , y(xn )) .
(40)
On expanding ln(I − ∆− ) into a Taylor series on the right-hand side of (40) we find that
1
1
1
19 4
y(xn+1 ) − y(xn ) ≈ h I − ∆− − ∆2− − ∆3− −
∆ − · · · f (xn , y(xn ))
2
12
24
720 −
(41)
and
1
5
3
251 4
y(xn+1 ) − y(xn ) ≈ h I + ∆− + ∆2− + ∆3− +
∆ + · · · f (xn , y(xn )) .
2
12
8
720 −
(42)
Successive truncation of (41) yields the family of Adams–Moulton methods, while similar
successive truncation of (42) gives rise to the family of Adams–Bashforth methods.
Next, we turn our attention to the analysis of linear multi-step methods and introduce
the concepts of stability, consistency and convergence.
3.2
Zero-stability
As is clear from (37) we need k starting values, y0 , . . . , yk−1 , before we can apply a linear kstep method to the initial value problem (1–2): of these, y0 is given by the initial condition
(2), but the others, y1 , . . . , yk−1 , have to be computed by other means: say, by using a
suitable Runge–Kutta method. At any rate, the starting values will contain numerical
errors and it is important to know how these will affect further approximations yn , n ≥ k,
which are calculated by means of (37). Thus, we wish to consider the ‘stability’ of the
numerical method with respect to ‘small perturbations’ in the starting conditions.
Definition 4 A linear k-step method for the ordinary differential equation y ′ = f (x, y)
is said to be zero-stable if there exists a constant K such that, for any two sequences
(yn ) and (ˆ
yn ) which have been generated by the same formulae but different initial data
y0 , y1 , . . . , yk−1 and yˆ0 , yˆ1 , . . . , yˆk−1 , respectively, we have
|yn − yˆn | ≤ K max{|y0 − yˆ0 |, |y1 − yˆ1 |, . . . , |yk−1 − yˆk−1 |}
for xn ≤ XM , and as h tends to 0.
24
(43)
We shall prove later on that whether or not a method is zero-stable can be determined
by merely considering its behaviour when applied to the trivial differential equation y ′ = 0,
corresponding to (1) with f (x, y) ≡ 0; it is for this reason that the kind of stability
expressed in Definition 4 is called zero stability. While Definition 4 is expressive in the
sense that it conforms with the intuitive notion of stability whereby “small perturbations
at input give rise to small perturbations at output”, it would be a very tedious exercise
to verify the zero-stability of a linear multi-step method using Definition 4 only; thus we
shall next formulate an algebraic equivalent of zero-stability, known as the root condition,
which will simplify this task. Before doing so we introduce some notation.
Given the linear k-step method (37) we consider its first and second characteristic
polynomial, respectively
k
X
ρ(z) =
αj z j ,
j=0
k
X
σ(z) =
βj z j ,
j=0
where, as before, we assume that
α20 + β02 6= 0 .
αk 6= 0 ,
Now we are ready to state the main result of this section.
Theorem 6 A linear multi-step method is zero-stable for any ordinary differential equation of the form (1) where f satisfies the Lipschitz condition (3), if and only if its first
characteristic polynomial has zeros inside the closed unit disc, with any which lie on the
unit circle being simple.
The algebraic stability condition contained in this theorem, namely that the roots of
the first characteristic polynomial lie in the closed unit disc and those on the unit circle
are simple, is often called the root condition.
Proof: Necessity. Consider the linear k-step method, applied to y ′ = 0:
αk yn+k + αk−1 yn+k−1 + · · · + α1 yn+1 + α0 yn = 0 .
(44)
The general solution of this kth order linear difference equation has the form
yn =
X
ps (n)zsn ,
(45)
s
where zs is a zero of the first characteristic polynomial ρ(z) and the polynomial ps (·)
has degree one less than the multiplicity of the zero. Clearly, if |zs | > 1 then there are
starting values for which the corresponding solutions grow like |zs |n and if |zs | = 1 and
its multiplicity is ms > 1 then there are solutions growing like nms −1 . In either case there
are solutions that grow unbounded as n → ∞, i.e. as h → 0 with nh fixed. Considering
starting data y0 , y1 , . . . , yk−1 which give rise to such an unbounded solution (yn ), and
starting data yˆ0 = yˆ1 = · · · = yˆk−1 = 0 for which the corresponding solution of (44) is (ˆ
yn )
with yˆn = 0 for all n, we see that (43) cannot hold. To summarise, if the root condition
is violated then the method is not zero-stable.
25
Sufficiency. The proof that the root condition is sufficient for zero-stability is long and
technical, and will be omitted here. For details, see, for example, K.W. Morton, Numerical
Solution of Ordinary Differential Equations, Oxford University Computing Laboratory,
1987, or P. Henrici, Discrete Variable Methods in Ordinary Differential Equations, Wiley,
New York, 1962. ⋄
Example 4 We shall consider the methods from Example 3.
a) The explicit and implicit Euler methods have first characteristic polynomial ρ(z) =
z − 1 with simple root z = 1, so both methods are zero-stable. The same is true of
the trapezium method.
b) The Adams–Bashforth and Adams–Moulton methods considered in Example 3 have
the same first characteristic polynomial, ρ(z) = z 3 (z −1), and therefore both methods
are zero-stable.
c) The three-step (sixth order accurate) linear multi-step method
11yn+3 + 27yn+2 − 27yn+1 − 11yn = 3h[fn+3 + 9fn+2 + 9fn+1 + fn ]
is not zero-stable. Indeed, the associated first characteristic polynomial ρ(z) = 11z 3 +
27z 2 − 27z − 11 has roots at z1 = 1, z2 ≈ −0.3189, z3 ≈ −3.1356, so |z3 | > 1.
3.3
Consistency
In this section we consider the accuracy of the linear k-step method (37). For this purpose, as in the case of one-step methods, we introduce the notion of truncation error.
Thus, suppose that y(x) is a solution of the ordinary differential equation (1). Then the
truncation error of (37) is defined as follows:
Tn =
Pk
j=0 [αj y(xn+j ) − hβj y
P
h kj=0 βj
′ (x
n+j )]
.
(46)
Of course, the definition requires implicitly that σ(1) = kj=0 βj 6= 0. Again, as in the
case of one-step methods, the truncation error can be thought of as the residual that is
obtained by inserting the solution of the differential equation into the formula (37) and
P
scaling this residual appropriately (in this case dividing through by h kj=0 βj ) so that Tn
resembles y ′ − f (x, y(x)).
P
Definition 5 The numerical scheme (37) is said to be consistent with the differential
equation (1) if the truncation error defined by (46) is such that for any ǫ > 0 there exists
h(ǫ) for which
|Tn | < ǫ for 0 < h < h(ǫ)
and any (k + 1) points (xn , y(xn )), . . . , (xn+k , y(xn+k )) on any solution curve in R of the
initial value problem (1–2).
26
Now let us suppose that the solution to the differential equation is sufficiently smooth,
and let us expand y(xn+j ) and y ′ (xn+j ) into a Taylor series about the point xn and
substitute these expansions into the numerator in (46) to obtain
Tn =
1
[C0 y(xn ) + C1 hy ′ (xn ) + C2 h2 y ′′ (xn ) + · · · ]
hσ(1)
(47)
where
C0
=
k
X
αj ,
j=0
C1
=
k
X
j=1
C2
=
=
Cq
k
X
2!
αj −
k
X
jq
αj −
j=1
q!
βj ,
j=0
k
X
j2
j=1
etc.
jαj −
k
X
jβj ,
j=1
k
X
j q−1
βj .
(q − 1)!
j=1
For consistency we need that Tn → 0 as h → 0 and this requires that C0 = 0 and C1 = 0;
in terms of the characteristic polynomials this consistency requirement can be restated in
compact form as
ρ(1) = 0 and ρ′ (1) = σ(1) 6= 0 .
Let us observe that, according to this condition, if a linear multi-step method is consistent
then it has a simple root on the unit circle at z = 1; thus the root condition is not violated
by this zero.
Definition 6 The numerical method (37) is said to have order of accuracy p if p is
the largest positive integer such that, for any sufficiently smooth solution curve in R of the
initial value problem (1–2), there exist constants K and h0 such that
|Tn | ≤ Khp
for 0 < h ≤ h0
for any (k + 1) points (xn , y(xn )), . . . , (xn+k , y(xn+k )) on the solution curve.
Thus we deduce from (47) that the method is of order of accuracy p if and only if
C0 = C1 = · · · = Cp = 0 and
In this case,
Tn =
Cp+1 6= 0 .
Cp+1 p (p+1)
h y
(xn ) + O(hp+1 ) ;
σ(1)
the number Cp+1 (6= 0) is called the error constant of the method.
Exercise 2 Construct an implicit linear two-step method of maximum order, containing
one free parameter. Determine the order and the error constant of the method.
27
Solution: Taking α0 = a as parameter, the method has the form
yn+2 + α1 yn+1 + ayn = h(β2 fn+2 + β1 fn+1 + β0 fn ) ,
with α2 = 1, α0 = a, β2 6= 0. We have to determine four unknowns: α1 , β2 , β1 , β0 , so we require
four equations; these will be arrived at by demanding that the constants
C0
C1
Cq
= α0 + α1 + α2 ,
= α1 + 2 − (β0 + β1 + β2 ) ,
1
1
(α1 + 2q α2 ) −
(β1 + 2q−1 β2 ) ,
=
q!
(q − 1)!
q = 2, 3 ,
appearing in (47) are all equal to zero, given that we wish to maximise the order of the method.
Thus,
C0
C1
=
=
C2
=
C3
=
a + α1 + 1 = 0 ,
α1 + 2 − (β0 + β1 + β2 ) = 0 ,
1
(α1 + 4) − (β1 + 2β2 ) = 0 ,
2!
1
1
(α1 + 8) − (β1 + 4β2 ) = 0 .
3!
2!
Hence,
α1 = −1 − a ,
1
β0 = − (1 + 5a) ,
12
β1 =
2
(1 − a) ,
3
β2 =
1
(5 + a) ,
12
and the resulting method is
yn+2 − (1 + a)yn+1 + ayn =
1
h [(5 + a)fn+2 + 8(1 − a)fn+1 − (1 + 5a)fn ] .
12
(48)
Further,
C4
=
C5
=
1
(α1 + 16) −
4!
1
(α1 + 32) −
5!
1
1
(β1 + 8β2 ) = − (1 + a) ,
3!
4!
1
1
(β1 + 16β2 ) = −
(17 + 13a) .
4!
3 · 5!
If a 6= −1 then C4 6= 0, and the method (48) is third order accurate. If, on the other hand, a = −1,
then C4 = 0 and C5 6= 0 and the method (48) becomes Simpson’s rule method – a fourth-order
accurate two-step method. The error constant is:
⋄
C4
=
C5
=
1
(1 + a) ,
4!
4
−
,
3 · 5!
−
a 6= −1 ,
(49)
a = −1 .
(50)
Exercise 3 Determine all values of the real parameter b, b 6= 0, for which the linear
multi-step method
yn+3 + (2b − 3)(yn+2 − yn+1 ) − yn = hb(fn+2 + fn+1 )
is zero-stable. Show that there exists a value of b for which the order of the method is 4.
Is the method convergent for this value of b? Show further that if the method is zero-stable
than its order cannot exceed 2.
28
Solution: According to the root condition, this linear multi-step method is zero-stable if and
only if all roots of its first characteristic polynomial
ρ(z) = z 3 + (2b − 3)(z 2 − z) − 1
belong to the closed unit disc, and those on the unit circle are simple.
Clearly, ρ(1) = 0; upon dividing ρ(z) by z − 1 we see that ρ(z) can be written in the following
factorised form:
ρ(z) = (z − 1) z 2 − 2(1 − b)z + 1 ≡ (z − 1)ρ1 (z) .
Thus the method is zero-stable if and only if all roots of the polynomial ρ1 (z) belong to the closed
unit disc, and those on the unit circle are simple and differ from 1. Suppose that the method is
zero-stable. Then, it follows that b 6= 0 and b 6= 2, since these values of b correspond to double
roots of ρ1 (z) on the unit circle, respectively, z = 1 and z = −1. Since the product of the two roots
of ρ1 (z) is equal to 1 and neither of them is equal to ±1, it follows that they are strictly complex;
hence the discriminant of the quadratic polynomial ρ1 (z) is negative. Namely,
4(1 − b)2 − 4 < 0 .
In other words, b ∈ (0, 2).
Conversely, suppose that b ∈ (0, 2). Then the roots of ρ(z) are
p
z1 = 1,
z2/3 = 1 − b + ı 1 − (b − 1)2 .
Since |z2/3 | = 1, z2/3 6= 1 and z2 6= z3 , all roots of ρ(z) lie on the unit circle and they are simple.
Hence the method is zero-stable.
To summarise, the method is zero-stable if and only if b ∈ (0, 2).
In order to analyse the order of accuracy of the method we note that upon Taylor series
expansion its truncation error can be written in the form
1
b
1
1
Tn =
1−
h2 y ′′′ (xn ) + (6 − b)h3 y IV (xn ) +
(150 − 23b)h4 y V (xn ) + O(h5 ) .
2b
6
4
120
If b = 6, then Tn = O(h4 ) and so the method is of order 4. As b = 6 does not belong to the
interval (0, 2), we deduce that the method is not zero-stable for b = 6.
Since zero-stability requires b ∈ (0, 2), in which case 1 − 6b 6= 0, it follows that if the method is
zero-stable then Tn = O(h2 ). ⋄
3.4
Convergence
The concepts of zero-stability and consistency are of great theoretical importance. However, what matters most from the practical point of view is that the numerically computed
approximations yn at the mesh-points xn , n = 0, . . . , N , are close to those of the analytical solution y(xn ) at these point, and that the global error en = y(xn ) − yn between
the numerical approximation yn and the exact solution-value y(xn ) decays when the step
size h is reduced. In order to formalise the desired behaviour, we introduce the following
definition.
Definition 7 The linear multistep method (37) is said to be convergent if, for all initial
value problems (1–2) subject to the hypotheses of Theorem 1, we have that
lim
h→0
yn = y(x)
(51)
nh=x−x0
holds for all x ∈ [x0 , XM ] and for all solutions {yn }N
n=0 of the difference equation (37)
with consistent starting conditions, i.e. with starting conditions ys = ηs (h), s =
0, 1, . . . , k − 1, for which limh→0 ηs (h) = y0 , s = 0, 1, . . . , k − 1.
29
We emphasize here that Definition 7 requires that (51) hold not only for those sequences
{yn }N
n=0 which have been generated from (37) using exact starting values ys = y(xs ),
s = 0, 1, . . . , k − 1, but also for all sequences {yn }N
n=0 whose starting values ηs (h) tend to
the correct value, y0 , as the h → 0. This assumption is made because in practice exact
starting values are usually not available and have to be computed numerically.
In the remainder of this section we shall investigate the interplay between zero-stability,
consistency and convergence; the section culminates in Dahlquist’s Equivalence Theorem
which, under some technical assumptions, states that for a consistent linear multi-step
method zero-stability is necessary and sufficient for convergence.
3.4.1
Necessary conditions for convergence
In this section we show that both zero-stability and consistency are necessary for convergence.
Theorem 7 A necessary condition for the convergence of the linear multi-step method
(37) is that it be zero-stable.
Proof: Let us suppose that the linear multi-step method (37) is convergent; we wish to
show that it is then zero-stable.
We consider the initial value problem y ′ = 0, y(0) = 0, on the interval [0, XM ], XM > 0,
whose solution is, trivially, y(x) ≡ 0. Applying (37) to this problem yields the difference
equation
αk yn+k + αk−1 yn+k−1 + · · · + α0 yn = 0 .
(52)
Since the method is assumed to be convergent, for any x > 0, we have that
lim yn = 0 ,
h→0
(53)
nh=x
for all solutions of (52) satisfying ys = ηs (h), s = 0, . . . , k − 1, where
lim ηs (h) = 0 ,
s = 0, 1, . . . k − 1 .
h→0
(54)
Let z = reiφ , be a root of the first characteristic polynomial ρ(z); r ≥ 0, 0 ≤ φ < 2π. It is
an easy matter to verify then that the numbers
yn = hr n cos nφ
define a solution to (52) satisfying (54). If φ 6= 0 and φ 6= π, then
yn2 − yn+1 yn−1
= h2 r 2n .
sin2 φ
Since the left-hand side of this identity converges to 0 as h → 0, n → ∞, nh = x, the
same must be true of the right-hand side; therefore,
lim
n→∞
2
x
n
30
r 2n = 0 .
This implies that r ≤ 1. In other words, we have proved that any root of the first
characteristic polynomial of (37) lies in the closed unit disc.
Next we prove that any root of the first characteristic polynomial of (37) that lies on
the unit circle must be simple. Assume, instead, that z = reiφ , is a multiple root of ρ(z),
with |z| = 1 (and therefore r = 1) and 0 ≤ φ < 2π. We shall prove below that this
contradicts our assumption that the method (52) is convergent. It is easy to check that
the numbers
yn = h1/2 nr n cos(nφ)
(55)
define a solution to (52) which satisfies (54), for
|ηs (h)| = |ys | ≤ h1/2 s ≤ h1/2 (k − 1) ,
s = 0, . . . k − 1 .
If φ = 0 or φ = π, it follows from (55) with h = x/n that
|yn | = x1/2 n1/2 r n .
(56)
Since, by assumption, |z| = 1 (and therefore r = 1), we deduce from (56) that limn→∞ |yn | =
∞, which contradicts (53).
If, on the other hand, φ 6= 0 and φ 6= π, then
zn2 − zn+1 zn−1
= r 2n ,
2
sin φ
(57)
where zn = n−1 h−1/2 yn = h1/2 x−1 yn . Since, by (53), limn→∞ zn = 0, it follows that
the left-hand side of (57) converges to 0 as n → ∞. But then the same must be true of
the right-hand side of (57); however, the right-hand side of (57) cannot converge to 0 as
n → ∞, since |r| = 1 (and hence r = 1). Thus, again, we have reached a contradiction.
To summarise, we have proved that all roots of the first characteristic polynomial ρ(z)
of the linear multi-step method (37) lie in the unit disc |z| ≤ 1, and those which belong
to the unit circle |z| = 1 are simple. By virtue of Theorem 6 the linear multi-step method
is zero-stable. ⋄
Theorem 8 A necessary condition for the convergence of the linear multi-step method
(37) is that it be consistent.
Proof: Let us suppose that the linear multi-step method (37) is convergent; we wish to
show that it is then consistent.
Let us first show that C0 = 0. We consider the initial value problem y ′ = 0, y(0) = 1,
on the interval [0, XM ], XM > 0, whose solution is, trivially, y(x) ≡ 1. Applying (37) to
this problem yields the difference equation
αk yn+k + αk−1 yn+k−1 + · · · + α0 yn = 0 .
(58)
We supply “exact” starting values for the numerical method; namely, we choose ys = 1,
s = 0, . . . , k − 1. Given that by hypothesis the method is convergent, we deduce that
lim yn = 1 .
h→0
nh=x
31
(59)
Since in the present case yn is independent of the choice of h, (59) is equivalent to saying
that
lim yn = 1 .
(60)
n→∞
Passing to the limit n → ∞ in (58), we deduce that
αk + αk−1 + · · · + α0 = 0 .
(61)
Recalling the definition of C0 , (61) is equivalent to C0 = 0 (i.e. ρ(1) = 0).
In order to show that C1 = 0, we now consider the initial value problem y ′ = 1,
y(0) = 0, on the interval [0, XM ], XM > 0, whose solution is y(x) = x. The difference
equation (37) now becomes
αk yn+k + αk−1 yn+k−1 + · · · + α0 yn = h(βk + βk−1 + · · · + β0 ) ,
(62)
where XM − x0 = XM − 0 = N h and 1 ≤ n ≤ N − k. For a convergent method every
solution of (62) satisfying
lim ηs (h) = 0 ,
h→0
s = 0, 1, . . . k − 1 ,
(63)
where ys = ηs (h), s = 0, 1, . . . , k − 1, must also satisfy
lim yn = x .
h→0
(64)
nh=x
Since according to the previous theorem zero-stability is necessary for convergence, we
may take it for granted that the first characteristic polynomial ρ(z) of the method does
not have a multiple root on the unit circle |z| = 1; therefore
ρ′ (1) = kαk + · · · + 2α2 + α1 6= 0 .
Let the sequence {yn }N
n=0 be defined by yn = Knh, where
K=
βk + · · · + β1 + β0
;
kαk + · · · + 2α2 + α1
(65)
this sequence clearly satisfies (63) and is the solution of (62). Furthermore, (64) implies
that
x = y(x) = lim yn = lim Knh = Kx ,
h→0
h→0
nh=x
nh=x
and therefore K = 1. Hence, from (65),
C1 = (kαk + · · · + 2α2 + α1 ) − (βk + · · · + β1 + β0 ) = 0 ;
equivalently, ρ′ (1) = σ(1) . ⋄
32
3.4.2
Sufficient conditions for convergence
We begin by establishing some preliminary results.
Lemma 1 Suppose that all roots of the polynomial ρ(z) = αk z k + · · · + α1 z + α0 lie in the
closed unit disk |z| ≤ 1 and those which lie on the unit circle |z| = 1 are simple. Assume
further that the numbers γl , l = 0, 1, 2, . . ., are defined by
1
= γ0 + γ1 z + γ2 z 2 + · · · .
αk + · · · + α1 z k−1 + α0 z k
Then, Γ ≡ supl≥0 |γl | < ∞ .
Proof: Let us define ρˆ(z) = z k ρ(1/z) and note that, by virtue of our assumptions about
the roots of ρ(z), the function 1/ˆ
ρ(z) is holomorphic in the open unit disc |z| < 1. As the
roots z1 , z2 , . . . , zm of ρ(z) on |z| = 1 are simple, the same is true of the poles of 1/ˆ
ρ(z),
and there exist constants A1 , . . . , Am such that the function
f (z) =
A1
1
Am
−
− ··· −
ρˆ(z) z − z1
z − z1m
1
(66)
is holomorphic for |z| < 1 and |f (z)| ≤ M for all |z| ≤ 1. Thus by Cauchy’s estimate8 the
coefficients of the Taylor expansion of f at z = 0 also form a bounded sequence. As
−
∞
X
Ai
=
A
zil z l ,
i
z − z1i
l=0
i = 1, . . . , m ,
and |zi | ≤ 1, we deduce from (66) that the coefficients in the Taylor series expansion of
1/ˆ
ρ(z) form a bounded sequence, which completes the proof. ⋄
Now we shall apply Lemma 1 to estimate the solution of the linear difference equation
αk em+k + αk−1 em+k−1 + · · · + α0 e0 = h(βk,m em+k + βk−1,m em+k−1 + · · · + β0,m em ) + λm .
(67)
The result is stated in the next Lemma.
Lemma 2 Suppose that all roots of the polynomial ρ(z) = αk z k + · · · + α1 z + α0 lie in the
closed unit disk |z| ≤ 1 and those which lie on the unit circle |z| = 1 are simple. Let B ∗
and Λ denote nonnegative constants and β a positive constant such that
|βk,n | + |βk−1,n | + · · · + |β0,n | ≤ B ∗ ,
|βk,n | ≤ β ,
|λn | ≤ Λ ,
n = 0, 1, . . . , N ,
and let 0 ≤ h < |αk |β −1 . Then every solution of (67) for which
|es | ≤ E ,
s = 0, 1, . . . , k − 1 ,
8
Theorem (Cauchy’s Estimate) If f is a holomorphic function in the open disc D(a, R), centre a
and radius R, and |f (z)| ≤ M for all z ∈ D(a, R), then |f (n) (a)| ≤ M (n!/Rn ), n = 0, 1, 2, . . . . [For a proof
of this result see, for example, Walter Rudin: Real and Complex Analysis. 3rd edition. McGraw-Hill, New
York, 1986.]
33
satisfies
|en | ≤ K ∗ exp(nhL∗ ) ,
n = 0, 1, . . . , N ,
where
L∗ = Γ∗ B ∗ ,
K ∗ = Γ∗ (N Λ + AEk) ,
Γ∗ = Γ/(1 − h|αk |−1 β) ,
Γ is as in Lemma 1, and
A = |αk | + |αk−1 | + · · · + |α0 | .
Proof: For a fixed k we consider the numbers γl , l = 0, 1, . . . , n − k, defined in Lemma 1.
After multiplying both sides of the equation (67) for m = n − k − l by γl , l = 0, 1, . . . , n − k
and summing the resulting equations, on denoting by Sn the sum, we find by manipulating
the left-hand side in the sum that
Sn = (αk en + αk−1 en−1 + · · · + α0 en−k )γ0
+(αk en−1 + αk−1 en−2 + · · · + α0 en−k−1 )γ1 + · · ·
+(αk ek + αk−1 ek−1 + · · · + α0 e0 )γn−k
= αk γ0 en + (αk γ1 + αk−1 γ0 )en−1 + · · ·
+(αk γn−k + αk−1 γn−k−1 + · · · + α0 γn−2k )ek
+(αk−1 γn−k + · · · + α0 γn−2k+1 )ek−1 + · · ·
+α0 γn−k e0 .
Defining γl = 0 for l < 0 and noting that
αk γl + αk−1 γl−1 + · · · + α0 γl−k =
(
1 for l = 0
0 for l > 0
we have that
Sn = en + (αk−1 γn−k + · · · + α0 γn−2k+1 )ek−1 + · · · + α0 γn−k e0 .
By manipulating similarly the right-hand side in the sum, we find that
en + (αk−1 γn−k + · · · + α0 γn−2k+1 )ek−1 + · · · + α0 γn−k e0
= h [βk,n−k γ0 en + (βk−1,n−k γ0 + βk,n−k−1 γ1 )en−1 + · · ·
+ (β0,n−k γ0 + · · · + βk,n−2k γk )en−k + · · · + β0,0 γn−k e0 ]
+(λn−k γ0 + λn−k−1 γ1 + · · · + λ0 γn−k ) .
Thus, by our assumptions and noting that by (68) γ0 = α−1
k ,
∗
|en | ≤ hβ|α−1
k | |en | + hΓB
n−1
X
m=0
|em | + N ΓΛ + AΓEk .
Hence,
∗
(1 − hβ|α−1
k |)|en | ≤ hΓB
n−1
X
m=0
34
|em | + N ΓΛ + AΓEk .
(68)
Recalling the definitions of L∗ and K ∗ we can rewrite the last inequality as follows:
|en | ≤ K ∗ + hL∗
n−1
X
|em | ,
m=0
n = 0, 1, . . . , N .
(69)
The final estimate is deduced from (69) by induction. First, we note that by virtue of
(68), AΓ ≥ 1. Consequently, K ∗ ≥ ΓAEk ≥ Ek ≥ E. Now, letting n = 1 in (69),
|e1 | ≤ K ∗ + hL∗ |e0 | ≤ K ∗ + hL∗ E ≤ K ∗ (1 + hL∗ ) .
Repeating this argument, we find that
|em | ≤ K ∗ (1 + hL∗ )m ,
m = 0, . . . , k − 1 .
Now suppose that this inequality has already been shown to hold for m = 0, 1, . . . , n − 1,
where n ≥ k; we shall prove that it then also holds for m = n, which will complete the
induction. Indeed, we have from (69) that
|en | ≤ K ∗ + hL∗ K ∗
(1 + hL∗ )n − 1
= K ∗ (1 + hL∗ )n .
hL∗
(70)
∗
Further, as 1 + hL∗ ≤ ehL we have from (70) that
|en | ≤ K ∗ ehL
∗n
,
n = 0, 1, . . . , N .
(71)
That completes the proof of the lemma. We remark that the implication (69) ⇒ (71) is
usually referred to as the Discrete Gronwall Lemma. ⋄
Using Lemma 2 we can now show that zero-stability and consistency, which have been
shown to be necessary are also sufficient conditions for convergence.
Theorem 9 For a linear multi-step method that is consistent with the ordinary differential equation (1) where f is assumed to satisfy a Lipschitz condition, and starting with
consistent starting conditions, zero-stability is sufficient for convergence.
Proof: Let us define
δ = δ(h) =
max |ηs (h) − y(a + sh)| ;
0≤s≤k−1
given that the starting conditions ys = ηs (h), s = 0, . . . , k−1, are assumed to be consistent,
we have that limh→0 δ(h) = 0. We have to prove that
lim
n→∞
yn = y(x)
nh=x−x0
for all x in the interval [x0 , XM ]. We begin the proof by estimating the truncation error
of (37):


Tn =
k
1 X
αj y(xn+j ) − hβj y ′ (xn+j )  .
hσ(1) j=0
35
(72)
As y ′ ∈ C[x0 , XM ], it makes sense to define, for ǫ ≥ 0, the function
χ(ǫ) =
max
|x∗ −x|≤ǫ
x, x∗ ∈[x0 ,XM ]
|y ′ (x∗ ) − y ′ (x)| .
For s = 0, 1, . . . , k − 1, we can then write
y ′ (xn+s ) = y ′ (xn ) + θs χ(sh) ,
where |θs | ≤ 1. Further, by the Mean Value Theorem, there exists ξs ∈ (xn , xn+s ) such
that
y(xn+s ) = y(xn ) + shy ′ (ξs ) .
Thus,
y(xm+s ) = y(xm ) + sh y ′ (xm ) + θs′ χ(sh) ,
where |θs′ | ≤ 1.
Now we can write
|σ(1)Tn | ≤ h−1 (α1 + α2 + · · · + αk )y(xn ) + (α1 + 2α2 + · · · + kαk )y ′ (xn )
− (β0 + β1 + · · · + βk )y ′ (xn )
+(|α1 | + 2|α2 | + · · · + k|αk |)|χ(kh)| + (|β0 | + |β1 | + · · · + |βk |)|χ(kh)| .
Since the method has been assumed consistent, the first, second and third terms on the
right-had side cancel, giving
|σ(1)Tn | ≤ (|α1 | + 2|α2 | + · · · + k|αk |)|χ(kh)| + (|β0 | + |β1 | + · · · + |βk |)|χ(kh)| .
Thus,
|σ(1)Tn | ≤ Kχ(kh) .
(73)
where
K = |α1 | + 2|α2 | + · · · + k|αk | + |β0 | + |β1 | + · · · + |βk | .
Comparing (37) with (72), we conclude that the global error em = y(xm ) − ym satisfies
αk em+k + · · · + α0 e0 = h (βk gm+k em+k + · · · + β0 gm em ) + σ(1)Tn h ,
where
gm =
(
[f (xm , y(xm )) − f (xm , ym )]/em , em =
6 0
0,
em = 0 .
By virtue of (73), we then have that
αk em+k + · · · + α0 e0 = h (βk gm+k em+k + · · · + β0 gm em ) + θKχ(kh)h .
As f is assumed to satisfy the Lipschitz condition (3) we have that |gm | ≤ L, m = 0, 1, . . ..
On applying Lemma 2 with E = δ(h), Λ = Kχ(kh)h, N = (XM − x0 )/h, B ∗ = BL, where
B = |β0 | + |β1 | + · · · + |βk |, we find that
|en | ≤ Γ∗ [Akδ(h) + (XM − x0 )Kχ(kh)] exp[(xn − x0 )LΓ∗ B] ,
36
(74)
where
A = |α0 | + |α1 | + · · · + |αk | ,
Γ∗ =
Γ
.
1 − h|α−1
k βk |L
Now, y ′ is a continuous function on the closed interval [x0 , XM ]; therefore it is uniformly
continuous on [x0 , XM ]. Thus, χ(kh) → 0 as h → 0; also, by virtue of the assumed
consistency of the starting values, δ(h) → 0 as h → 0. Passing to the limit h → 0 in (74),
we deduce that
lim |en | = 0 ;
n→∞
x−x0 =nh
equivalently,
lim
n→∞
x−x0 =nh
|y(x) − yn | = 0
so the method is convergent. ⋄
On combining Theorems 7, 8 and 9, we arrive at the following important result.
Theorem 10 (Dahlquist) For a linear multi-step method that is consistent with the
ordinary differential equation (1) where f is assumed to satisfy a Lipschitz condition,
and starting with consistent initial data, zero-stability is necessary and sufficient for convergence. Moreover if the solution y(x) has continuous derivative of order (p + 1) and
truncation error O(hp ), then the global error en = y(xn ) − yn is also O(hp ).
By virtue of Dahlquist’s theorem, if a linear multi-step method is not zero-stable its
global error cannot be made arbitrarily small by taking the mesh size h sufficiently small
for any sufficiently accurate initial data. In fact, if the root condition is violated then
there exists a solution to the linear multi-step method which will grow by an arbitrarily
large factor in a fixed interval of x, however accurate the starting conditions are. This
result highlights the importance of the concept of zero-stability and indicates its relevance
in practical computations.
3.5
Maximum order of a zero-stable linear multi-step method
Let us suppose that we have already chosen the coefficients αj , j = 0, . . . , k, of the k-step
method (37). The question we shall be concerned with in this section is how to choose
the coefficients βj , j = 0, . . . , k, so that the order of the resulting method (37) is as high
as possible.
In view of Theorem 10 we shall only be interested in consistent methods, so it is natural
to assume that the first and second characteristic polynomials ρ(z) and σ(z) associated
with (37) satisfy ρ(1) = 0, ρ′ (1) − σ(1) = 0, with σ(1) 6= 0 (the last condition being
required for the sake of correctness of the definition of the truncation error (46)).
Consider the function φ of the complex variable z, defined by
φ(z) =
ρ(z)
− σ(z) ;
log z
(75)
the function log z appearing in the denominator is made single-valued by cutting the
complex plane along the half-line ℜz ≤ 0. We begin our analysis with the following
fundamental lemma.
37
Lemma 3 Suppose that p is a positive integer. The linear multistep method (37), with
stability polynomials ρ(z) and σ(z), is of order of accuracy p if and only if the function
φ(z) defined by (75) has a zero of multiplicity p at z = 1.
Proof: Let us suppose that the k-step method (37) for the numerical approximation of
the initial value problem (1–2) is of order p. Then, for any sufficiently smooth function
f (x, y), the resulting solution to (1–2) yields a truncation error of the form:
Tn =
Cp+1 p (p+1)
h y
(xn ) + O(hp+1 ) ,
σ(1)
as h → 0, Cp+1 6= 0, xn = x0 + nh. In particular, for the initial value problem
y′ = y ,
we get
y(0) = 1 ,
i
enh h h
Cp+1 p
ρ(e ) − hσ(eh ) = enh
h + O(hp+1 ) ,
hσ(1)
σ(1)
6 0. Thus, the function
=
Tn ≡
as h → 0, Cp+1
(76)
i
1h h
ρ(e ) − hσ(eh )
h
is holomorphic in a neighbourhood of h = 0 and has a zero of order p at h = 0. The
function z = eh is a bijective mapping of a neighbourhood of h = 0 onto a neighbourhood
of z = 1. Therefore φ(z) is holomorphic in a neighbourhood of z = 1 and has a zero of
multiplicity p at z = 1.
Conversely, suppose that φ(z) has a zero of multiplicity p at z = 1. Then f (h) = φ(eh )
is a holomorphic function in the vicinity of h = 0 and has a zero of multiplicity p at h = 0.
Therefore,
f (h) =
g(h) =
k
X
(αj − hβj )ejh
j=0
has a zero of multiplicity (p + 1) at h = 0, implying that g(0) = g′ (0) = · · · = g(p) (0) = 0,
but g(p+1) (0) 6= 0. First,
g(0) = 0 =
k
X
αj = C0 .
j=0
Now, by successive differentiation of g with respect to h,
g′ (0) = 0 =
k
X
(jαj − βj ) = C1 ,
j=0
′′
g (0) = 0 =
k
X
(j 2 αj − 2jβj ) = 2C2 ,
j=0
g′′′ (0) = 0 =
k
X
(j 3 αj − 3j 2 βj ) = 6C3 ,
j=0
etc.
g(p) (0) = 0 =
k
X
(j p αj − pj p−1 βj ) = p! Cp .
j=0
38
We deduce that C0 = C1 = C2 = · · · = Cp = 0; since g(p+1) (0) 6= 0 we have that Cp+1 6= 0.
Consequently (37) is of order of accuracy p. ⋄
We shall use this lemma in the next theorem to supply a lower bound for the maximum
order of a linear multistep method with prescribed first stability polynomial ρ(z).
Theorem 11 Suppose that ρ(z) is a polynomial of degree k such that ρ(1) = 0 and ρ′ (1) 6=
0, and let kˆ be an integer such that 0 ≤ kˆ ≤ k. Then there exists a unique polynomial
σ(z) of degree kˆ such that ρ′ (1) − σ(1) = 0 and the order of the linear multi-step method
associated with ρ(z) and σ(z) is ≥ kˆ + 1.
Proof: Since the function ρ(z)/ log(z) is holomorphic in the neighbourhood of z = 1 it
can be expanded into a convergent Taylor series:
ρ(z)
= c0 + c1 (z − 1) + c2 (z − 1)2 + · · · .
log z
On multiplying both sides by log z and differentiating we deduce that c0 = ρ′ (1) (6= 0).
Let us define
ˆ
σ(z) = c0 + c1 (z − 1) + · · · + ckˆ (z − 1)k .
Clearly σ(1) = c0 = ρ′ (1) (6= 0). With this definition,
φ(z) =
ρ(z)
ˆ
k+1
− σ(z) = ck+1
+ ··· ,
ˆ (z − 1)
log z
and therefore φ(z) has a zero at z = 1 of multiplicity not less than kˆ + 1. By Lemma 3
the linear k-step method associated with ρ(z) and σ(z) is of order ≥ kˆ + 1.
The uniqueness of σ(z) possessing the desired properties follows from the uniqueness
of the Taylor series expansion of φ(z) about the point z = 1. ⋄
We note in connection with this theorem that for most methods of practical interest
either kˆ = k − 1 resulting in an explicit method or kˆ = k corresponding to an implicit
method. In the next example we shall encounter the latter situation.
Example 5 Consider a linear two-step method with ρ(z) = (z − 1)(z − λ). The method
will be zero-stable as long as λ ∈ [−1, 1). Consider the Taylor series expansion of the
function ρ(z)/ log(z) about the point z = 1:
ρ(z)
log z
=
(z − 1)(1 − λ + (z − 1))
log[1 + (z − 1)]
(
)−1
(
)
z − 1 (z − 1)2 (z − 1)3
+
−
+ O((z − 1)4 )
= [1 − λ + (z − 1)] × 1 −
2
3
4
z − 1 (z − 1)2 (z − 1)3
= [1 − λ + (z − 1)] × 1 +
−
+
+ O((z − 1)4 )
2
12
24
3−λ
5+λ
1+λ
= 1−λ+
(z − 1) +
(z − 1)2 −
(z − 1)3 + O((z − 1)4 ) .
2
12
24
A two-step method of maximum order is obtained by selecting
3−λ
5+λ
(z − 1) +
(z − 1)2
2
12
1 + 5λ 2 − 2λ
5+λ 2
= −
+
z+
z .
12
3
12
σ(z) = 1 − λ +
39
If λ 6= −1, the resulting method is of third order, with error constant
C4 = −
1+λ
,
24
whereas if λ = −1 the method is of fourth order.
In the former case the method is
2 − 2λ
1 + 5λ
5+λ
yn+2 − (1 + λ)yn+1 + λyn = h
fn+2 +
fn+1 −
fn
12
3
12
with λ a parameter contained in the interval (−1, 1). In the latter case, the method has
the form
h
yn+2 − yn = (fn+2 + 4fn+1 + fn ) ,
3
and is referred to as Simpson’s method.
By inspection, the linear k-step method (37) has 2k+2 coefficients: αj , βj , j = 0, . . . , k,
of which αk is taken to be 1 by normalisation. This leaves us with 2k + 1 free parameters
if the method is implicit and 2k free parameters if the method is explicit (given that in
the latter case βk is fixed to have value 0). According to (47), if the method is required to
have order p, p + 1 linear relationships C0 = 0, . . . , Cp = 0 involving αj , βj , j = 0, . . . , k,
must be satisfied. Thus, in the case of the implicit method, we can impose p + 1 = 2k + 1
linear constraints C0 = 0, . . . , C2k+1 = 0 to determine the unknown constants, yielding
a method of order p = 2k. Similarly, in the case of an explicit method, the highest order
we can expect is p = 2k − 1. Unfortunately, there is no guarantee that such methods will
be zero-stable. Indeed, in a paper published in 1956 Dahlquist proved that there is no
consistent, zero-stable k-step method which is of order > (k + 2). Therefore the maximum
orders 2k and 2k − 1 cannot be attained without violating the condition of zero-stability.
We formalise these facts in the next theorem.
Theorem 12 There is no zero-stable linear k-step method whose order exceeds k + 1 if k
is odd or k + 2 if k is even.
Proof: Consider a linear k-step method (37) with associated first and second stability
polynomials ρ and σ. Further, consider the transformation
ζ ∈ C 7→
ζ −1
∈C
ζ +1
which maps the open unit disc |ζ| < 1 of the ζ-plane onto the open half-plane ℜz < 0 of
the z-plane; the circle |ζ| = 1 is mapped onto the imaginary axis ℜz = 0, the point ζ = 1
onto z = 0, and the point ζ = −1 onto ζ = ∞.
It is clear that the functions r and s defined by
r(z) =
1−z
2
k
1+z
ρ
1−z
,
s(z) =
are in fact polynomials, deg(r) ≤ k and deg(s) ≤ k.
40
1−z
2
k 1+z
ρ
1−z
,
If ρ(ζ) has a root of multiplicity p, 0 ≤ p ≤ k, at ζ = ζ0 6= −1, then r(z) has a root
of the same multiplicity at z = (ζ0 − 1)/(ζ0 + 1); if ρ(ζ) has a root of multiplicity p ≥ 1,
0 ≤ p ≤ k, at ζ = −1, then r(z) is of degree k − p.
Since, by assumption, the method is zero-stable, ζ = 1 is a simple root of ρ(ζ); consequently, z = 0 is a simple root of r(z). Thus,
r(z) = a1 z + a2 z 2 + · · · + ak z k ,
a1 6= 0 ,
aj ∈ R .
It can be assumed, without loss of generality, that a1 > 0. Since by zero stability all roots
of ρ(ζ) are contained in the closed unit disc, we deduce that all roots of r(z) have real
parts ≤ 0. Therefore, all coefficients aj , j = 1, . . . , k, of r(z) are nonnegative.
Now let us consider the function
q(z) =
1−z
2
k
1+z
φ
1−z
=
1
log
1+z
1−z
r(z) − s(z) .
The function q(z) has a zero of multiplicity p at z = 0 if and only if φ(ζ) defined by (75)
has a zero of multiplicity p at ζ = 1; according to Lemma 3 this is equivalent to the linear
k-step method associated with ρ(ζ) and σ(ζ) having order p. Thus if the linear k-step
method associated with ρ(z) and σ(z) has order p then
s(z) = b0 + b1 z + b2 z 2 + · · · + bp−1 z p−1 ,
where
z
log
1+z
1−z
r(z)
= b0 + b1 z + b2 z 2 + · · · .
z
As the degree of s(z) is ≤ k, the existence of a consistent zero-stable k-step linear multistep
method of order p > k + 1 (or p > k + 2) now hinges on the possibility that
bk+1 = · · · = bp−1 = 0 ,
(or bk+2 = · · · = bp−1 = 0) .
Let us consider whether this is possible.
We denote by c0 , c1 , c2 , . . ., the coefficients in the Taylor series expansion of the function
z
log
namely,
z
log
1+z
1−z
1+z
1−z
,
= c0 + c2 z 2 + c4 z 4 + · · · .
Then, adopting the notational convention that aν = 0 for ν > k, we have that
b0
=
c0 a0 ,
b1
=
c0 a2 ,
etc.
b2ν
=
b2ν+1
=
c0 a2ν+1 + c2 a2ν−1 + · · · + c2ν a1 ,
c0 a2ν+2 + c2 a2ν + · · · + c2ν a2 ,
ν = 1, 2, . . . .
It is a straightforward matter to prove that c2ν < 0, ν = 1, 2, . . . (see also Lemma 5.4 on
page p.233 of Henrici’s book).
41
(i) If k is an odd number, then, since aν = 0 for ν > k, we have that
bk+1 = c2 ak + c4 ak−2 + · · · + ck+1 a1 .
Since a1 > 0 and no aν is negative, it follows that bk+1 < 0.
(ii) If k is an even number, then
bk+1 = c2 ak + c4 ak−2 + · · · + ck a2 .
Since c2ν < 0, ν = 1, 2, . . . , and aµ ≥ 0, µ = 2, 3, . . . , k, we deduce that bk+1 = 0 if
and only if a2 = a4 = · · · = ak = 0, i.e. when r(z) is an odd function of z. This,
together with the fact that all roots of r(z) have real part ≤ 0, implies that all roots
of r(z) mush have real part equal to zero. Consequently, all roots of ρ(ζ) lie on
|ζ| = 1. Since ak = 0, the degree of r(z) is k − 1, and therefore −1 is a (simple) root
of ρ(ζ).
As c2ν < 0, aµ ≥ 0 and a1 > 0, it follows that
bk+2 = c4 ak−1 + c6 ak−3 + · · · + ck+2 a1 < 0 ,
showing that bk+2 6= 0.
Thus if a k-step method is zero-stable and k is odd then bk+1 6= 0, whereas if k is even
then bk+2 . This proves that there is no zero-stable k-step method whose order exceeds
k + 1 if k is odd or k + 2 if k is even. ⋄
Definition 8 A zero-stable linear k-step method of order k + 2 is said to be an optimal
method.
According to the proof of the previous theorem, all roots of the first characteristic
polynomial ρ associated with an optimal linear multistep method have modulus 1.
Example 6 As k + 2 = 2k if an only if k = 2 and Simpson’s rule method is the zero-stable
linear 2-step method of maximum order, we deduce that Simpson’s rule method is the only
zero-stable linear multistep method which is both of maximum order (2k = 4) and optimal
(k + 2 = 4).
Optimal methods have certain disadvantages in terms of their stability properties; we
shall return to this question later on in the notes.
Linear k-step methods for which the first characteristic polynomial has the form
ρ(z) = z k − z k−1 are called Adams methods. Explicit Adams methods are referred to
as Adams–Bashforth methods, while implicit Adams methods are termed Adams–
Moulton methods. Linear k-step methods for which ρ(z) = z k − z k−2 are called
Nystr¨
om methods if explicit and Milne–Simpson methods if implicit. All these
methods are zero-stable.
42
3.6
Absolute stability of linear multistep methods
Up to now we have been concerned with the stability and accuracy properties of linear
multistep methods in the asymptotic limit of h → 0, n → ∞, nh fixed. However, it is of
practical significance to investigate the performance of methods in the case of h > 0 fixed
and n → ∞. Specifically, we would like to ensure that when applied to an initial value
problem whose solution decays to zero as x → ∞, the linear multistep method exhibits a
similar behaviour, for h > 0 fixed and xn = x0 + nh → ∞.
The canonical model problem with exponentially decaying solution is
y ′ = λy ,
where ℜλ < 0. Indeed,
x>0,
y(0) = y0 (6= 0) ,
(77)
y(x) = y0 eıxℑλ exℜλ ,
and therefore,
|y(x)| ≤ exp(−x|ℜλ|) ,
x≥0,
yielding limx→∞ y(x) = 0 . Thus, using the terminology introduced in the last paragraph
of Section 1, the solution is asymptotically stable.
In the rest of the section we shall assume, for simplicity, that λ is a negative real
number, but everything we shall say extends straightforwardly to the general case of λ
complex, ℜλ < 0.
Now consider the linear k-step method (37) and apply it to the model problem (77)
with λ real and negative. This yields the linear difference equation
k
X
j=0
(αj − hλβj ) yn+j = 0 .
Given that the general solution yn to this homogeneous difference equation can be expressed as a linear combination of powers of roots of the associated characteristic polynomial
¯ = ρ(z) − hσ(z)
¯
¯ = hλ) ,
π(z; h)
,
(h
(78)
it follows that yn will converge to zero for h > 0 fixed and n → ∞ if and only if all roots of
¯ defined by (78) is called the
π(z; h) have modulus < 1. The kth degree polynomial π(z; h)
stability polynomial of the linear k-step method with first and second characteristic
polynomials ρ(z) and σ(z), respectively. This motivates the following definition.
Definition 9 The linear multistep method (37) is called absolutely stable for a given
¯ if and only if for that h
¯ all the roots rs = rs (h)
¯ of the stability polynomial π(z, h)
¯ defined
h
by (78) satisfy |rs | < 1, s = 1, . . . , k. Otherwise, the method is said to be absolutely
unstable. An interval (α, β) of the real line is called the interval of absolute stability
¯ ∈ (α, β). If the method is absolutely unstable
if the method is absolutely stable for all h
¯ it is said to have no interval of absolute stability.
for all h,
Since for λ > 0 the solution of (77) exhibits exponential growth, it is reasonable
to expect that a consistent and zero-stable (and, therefore, convergent) linear multistep
method will have a similar behaviour for h > 0 sufficiently small, and will be therefore
¯ = λh. According to the next theorem, this is indeed the
absolutely unstable for small h
case.
43
Theorem 13 Every consistent and zero-stable linear multistep method is absolutely un¯
stable for small positive h.
Proof: Given that the method is consistent, there exists an integer p ≥ 1 such that
C0 = C1 = · · · = Cp = 0 and Cp+1 6= 0. Let us consider
¯ ¯
¯
¯
h
¯
π(eh ; h)
= ρ(eh ) − hσ(e
)
=
k h
X
j=0
=
k
X
j=0
=
k
X
j=0
=
k
X
¯
¯
¯ j ehj
αj ehj − hβ

αj

αj
∞ ¯ q
X
(hj)
q=0
αj +
j=0
k
X
− βj
q!
∞ ¯ q
X
(hj)
q=0
k
X
j=0
− βj
q!

αj

i
∞ ¯ q+1 q
X
h j
q=0

∞ ¯ q q−1
X
h j
(q − 1)!
q=1
∞ ¯ q
X
(hj)
q=1
q!

q!
− βj


∞ ¯ q q−1
X
h j
(q − 1)!
q=1



k
jq X
j q−1 
¯q 
=
αj +
h
αj −
βj
q! j=0 (q − 1)!
q=1
j=0
j=0
= C0 +
∞
X
∞
X
k
X
¯ q Cq
h
q=1
=
∞
X
¯ q = O(h
¯ p+1 ) .
Cq h
(79)
q=p+1
¯ can be written in the factorised
On the other hand, noting that the polynomial π(z; h)
form
¯ = (αk − hβ
¯ k )(z − r1 ) · · · (z − rk )
π(z, h)
¯ s = 1, . . . , k, signify the roots of π(.; h),
¯ we deduce that
where rs = rs (h),
¯ ¯
¯ k )(eh¯ − r1 (h))
¯ · · · (eh¯ − rk (h))
¯ .
π(eh ; h)
= (αk − hβ
(80)
¯ k → αk 6= 0 and rs (h)
¯ → ζs , s = 1, . . . , k, where ζs , s = 1, . . . , k,
¯ → 0, αk − hβ
As h
are the roots of the first stability polynomial ζ(z). Since, by assumption, the method is
consistent, 1 is a root of ζ(z); furthermore, by zero-stability 1 is a simple root of ζ(z). Let
us suppose, for the sake of definiteness that it is ζ1 that is equal to 1. Then, ζs 6= 1 for
s 6= 1 and therefore
¯
¯ = (1 − ζs ) 6= 0 ,
lim (eh − rs (h))
¯
h→0
s 6= 1 .
¯ ¯
¯ → 0 is
We conclude from (80) that the only factor of π(eh ; h)
that converges to 0 as h
¯
h
¯ (the other factors converge to nonzero constants). Now, by (79), π(e¯h ; h)
¯ =
e − r1 (h)
p+1
¯
O(h ), so it follows that
¯
¯ = O(h
¯ p+1 ) .
eh − r1 (h)
44
Thus we have shown that
¯
¯ = eh + O(h
¯ p+1 ) .
r1 (h)
This implies that
¯
¯ > 1 + 1h
r1 (h)
2
¯ That completes the proof. ⋄
for small positive h.
According to the definition adopted in the previous section, an optimal k-step method
is a zero-stable linear k-step method of order k + 2. We have also seen in the proof
of Theorem 12 that all roots of the first characteristic polynomial of an optimal k-step
method lie on the unit circle. By refining the proof of Theorem 13 it can be shown that
an optimal linear multistep method has no interval of absolute stability.
It also follows from Theorem 13 that whenever a consistent zero-stable linear multistep
method is used for the numerical solution of the initial value problem (1–2) where ∂f
∂y > 0,
the error of the method will increase as the computation proceeds.
3.7
General methods for locating the interval of absolute stability
In this section we shall describe two methods for identifying the endpoints of the interval
of absolute stability. The first of these is based on the Schur criterion, the second on the
Routh–Hurwitz criterion.
3.7.1
The Schur criterion
Consider the polynomial
φ(r) = ck r k + · · · + c1 r + c0 ,
ck 6= 0 ,
c0 6= 0 ,
with complex coefficients. The polynomial φ is said to be a Schur polynomial if each
of its roots rs satisfies |rs | < 1, s = 1, . . . , k.
Let us consider the polynomial
ˆ
φ(r)
= c¯0 r k + c¯1 r k−1 + · · · + c¯k−1 r + c¯k ,
where c¯j denotes the complex conjugate of cj , j = 1, . . . , k. Further, let us define
φ1 (r) =
i
1 hˆ
ˆ
φ(0)φ(r) − φ(0)φ(r)
.
r
Clearly φ1 has degree ≤ k − 1.
The following key result is stated without proof.
Theorem 14 (Schur’s Criterion) The polynomial φ is a Schur polynomial if and only
ˆ
if |φ(0)|
> |φ(0)| and φ1 is a Schur polynomial.
We illustrate Schur’s criterion by a simple example.
Exercise 4 Use Schur’s criterion to determine the interval of absolute stability of the
linear multistep method
h
yn+2 − yn = (fn+1 + 3fn ) .
2
45
Solution: The first and second characteristic polynomials of the method are
ρ(z) = z 2 − 1 ,
σ(z) =
1
(z + 3) .
2
Therefore the stability polynomial is
1¯
2
¯
¯
π(r; h) = ρ(r) − hσ(r) = r − hr − 1 +
2
3¯
h .
2
Now,
3¯ 2
¯
π
ˆ (r; h) = − 1 + h r −
2
¯ > |ˆ
¯ if and only if h
¯ ∈ (− 4 , 0). As
Clearly, |ˆ
π (0; h)|
π (0, h)|
3
1¯
hr + 1 .
2
ˆ = − 1 h(2
¯ + 3 h)(3r
¯
π1 (r, h)
+ 1)
2
2
has the unique root − 31 and is, therefore, a Schur polynomial, we deduce from Schur’s criterion
¯ is a Schur polynomial if and only if ¯h ∈ (− 4 , 0). Therefore the interval of absolute
that π(r; h)
3
stability is (− 43 , 0). ⋄
3.7.2
The Routh–Hurwitz criterion
Consider the mapping
r−1
r+1
of the open unit disc |r| < 1 of the complex r-plane to the open left half-plane ℜz < 0 of
the complex z-plane. The inverse of this mapping is given by
z=
r=
Under this transformation the function
1+z
.
1−z
¯ = ρ(r) − hσ(r)
¯
π(r, h)
becomes
ρ
1+z
1−z
¯
− hσ
1+z
1−z
.
On multiplying this by (1 − z)k , we obtain a polynomial of the form
a0 z k + a1 z k−1 + · · · + ak .
(81)
¯ lie inside the open unit disk |r| < 1 if and
The roots of the stability polynomial π(r, h)
only if the roots of the polynomial (81) lie in the open left half-plane ℜz < 0.
Theorem 15 (Routh–Hurwitz Criterion) The roots of (81) lie in the open left halfplane if and only if all the leading principal minors of the k × k matrix





Q=



a1 a3 a5
a0 a2 a4
0 a1 a3
0 a0 a2
··· ··· ···
0
0
0
···
···
···
···
···
···
a2k−1
a2k−2
a2k−3
a2k−4
···
ak









are positive and a0 > 0; we assume that aj = 0 if j > k. In particular,
46
a) for k = 2: a0 > 0, a1 > 0, a2 > 0,
b) for k = 3: a0 > 0, a1 > 0, a2 > 0, a3 > 0, a1 a2 − a3 a0 > 0,
c) for k = 4: a0 > 0, a1 > 0, a2 > 0, a3 > 0, a4 > 0, a1 a2 a3 − a0 a23 − a4 a21 > 0
represent the necessary and sufficient conditions for ensuring that all roots of (81) lie in
the open left half-plane.
We illustrate this result by a simple exercise.
Exercise 5 Use the Routh–Hurwitz criterion to determine the interval of absolute stability
of the linear multistep method from the previous exercise.
Solution: On applying the substitution
r=
1+z
1−z
in the stability polynomial
¯ − 1 + 3h
¯
¯ = r2 − 1 hr
π(r, h)
2
2
and multiplying the resulting function by (1 − z)2 , we get
"
2
#
1+z
1¯ 1 + z
3¯
2
(1 − z)
− h
− 1+ h
= a0 z 2 + a1 z + a2
1−z
2
1−z
2
with
¯,
a0 = − h
a1 = 4 + 3h ,
¯.
a2 = −2h
¯ ∈ (− 4 , 0);
Applying part a) of Theorem 15 we deduce that the method is zero-stable if and only if h
3
4
hence the interval of absolute stability is (− 3 , 0). ⋄
We conclude this section by listing the intervals of absolute stability (α, 0) of k-step
Adams–Bashforth and Adams–Moulton methods, for k = 1, 2, 3, 4. We shall also supply
the orders p∗ and p and error constants Cp∗ +1 and Cp+1 , respectively, of these methods.
The verification of the stated properties is left to the reader as exercise.
k-step Adams–Bashforth (explicit) methods:
(1) k = 1, p∗ = 1, Cp∗ +1 = 12 , α = −2 ,
y1 − y0 = hf0 ;
(2) k = 2, p∗ = 2, Cp∗ +1 =
5
12 ,
α = −1 ,
y2 − y1 =
h
(3f1 − f0 ) ;
2
6
(3) k = 3, p∗ = 3, Cp∗ +1 = 38 , α = − 11
,
y3 − y2 =
h
(23f2 − 16f1 + 5f0 ) ;
12
47
(4) k = 4, p∗ = 4, Cp∗ +1 =
251
720 ,
3
α = − 10
,
y4 − y3 =
h
(55f3 − 59f2 + 37f1 − 9f0 ) .
24
k-step Adams–Moulton (implicit) methods:
1
(1) k = 1, p = 2, Cp+1 = − 12
, α = −∞ ,
y1 − y0 =
h
(f1 + f0 ) ;
2
1
, α = −6 ,
(2) k = 2, p = 3, Cp+1 = − 24
y2 − y1 =
h
(5f2 + 8f1 − f0 ) ;
12
19
(3) k = 3, p = 4, Cp+1 = − 720
, α = −3 ,
y3 − y2 =
h
(9f3 + 19f2 − 5f1 + f0 ) ;
24
27
(4) k = 4, p = 5, Cp+1 = − 1440
, α = − 90
49 ,
y4 − y3 =
h
(251f4 + 646f3 − 264f2 + 106f1 − 19f0 ) .
720
We notice that the k-step Adams–Moulton (implicit) method has larger interval of absolute
stability and smaller error constant than the k-step Adams–Bashforth (explicit) method.
3.8
Predictor-corrector methods
Let us suppose that we wish to use the implicit linear k-step method
k
X
j=0
αj yn+j = h
k
X
βj fn+j ,
αk , βk 6= 0 .
j=0
Then, at each step we have to solve for yn+k the equation
αk yn+k − hβk f (xn+k , yn+k ) =
k−1
X
j=0
(hβj fn+j − αj yn+j ) .
If h < |αk |/(L|βk |) where L is the Lipschitz constant of f with respect to y (as in Picard’s Theorem 1), then this equation has a unique solution, yn+k ; moreover, yn+k can be
computed by means of the fixed-point iteration
[s+1]
αk yn+k +
k−1
X
[s]
αj yn+j = hβk f (xn+k , yn+k ) + h
j+0
k−1
X
j=0
48
βj fn+j ,
s = 0, 1, 2, . . . ,
[0]
with yn+k a suitably chosen starting value.
[s]
Theoretically, we would iterate until the iterates yn+k have converged (in practice, until
[s+1]
[s]
some stopping criterion such as |yn+k − yn+k | < ǫ is satisfied, where ǫ is some preassigned
tolerance). We would then regard the converged value as an acceptable approximation
yn+k to the unknown analytical solution-value y(xn+k ). This procedure is usually referred
to as correcting to convergence.
Unfortunately, in practice, such an approach is usually unacceptable due to the amount
[s]
of work involved: each step of the iteration involves an evaluation of f (xn+k , yn+k ) which
may be quite time-consuming. In order to keep to a minimum the number of times
[s]
[0]
f (xn+k , yn+k ) is evaluated, the initial guess yn+k must be chosen accurately. This is
[0]
achieved by evaluating yn+k by a separate explicit method called the predictor, and
taking this as the initial guess for the iteration based on the implicit method. The implicit
method is called the corrector.
For the sake of simplicity we shall suppose that the predictor and the corrector have the
same number of steps, say k, but in the case of the corrector we shall allow that both α0
and β0 vanish. Let the linear multistep method used as predictor have the characteristic
polynomials
ρ∗ (z) =
k
X
α∗j z j ,
α∗k = 1 ,
σ ∗ (z) =
j=0
k−1
X
βj∗ z j ,
(82)
j=0
and suppose that the corrector has characteristic polynomials
ρ(z) =
k
X
j
αj z ,
αk = 1 ,
σ(z) =
j=0
k
X
βj z j .
(83)
j=0
Suppose that m is a positive integer: it will denote the number of times the corrector is
applied; in practice m ≤ 2. If P indicates the application of the predictor, C a single
application of the corrector, and E an evaluation of f in terms of the known values of its
arguments, then P (EC)m E and P (EC)m denote the following procedures.
a) P (EC)m E
[0]
yn+k +
k−1
X
[m]
α∗j yn+j = h
j=0
k−1
X
[m]
βj∗ fn+j ,
j=0
[s]
[s]
fn+k = f (xn+k , yn+k ) ,
[s+1]
yn+k +
k−1
X
[m]
[s]
αj yn+j = hβk fn+k + h
j=0
k−1
X
j=0
[m]
[m]
fn+k = f (xn+k , yn+k ) ,
for n = 0, 1, 2, . . ..
b) P (EC)m
[0]
yn+k +
k−1
X
j=0
[m]
α∗j yn+j = h
k−1
X
[m−1]
βj∗ fn+j
j=0
49
,
[m]
βj fn+j ,
s = 0, . . . , m − 1,
[s]
[s]
fn+k = f (xn+k , yn+k ) ,
[s+1]
yn+k +
k−1
X
[m]
[s]
αj yn+j = hβk fn+k + h
j=0
k−1
X
[m−1]
βj fn+j
,
s = 0, . . . , m − 1 ,
j=0
for n = 0, 1, 2, . . ..
3.8.1
Absolute stability of predictor-corrector methods
Let us apply the predictor-corrector method P (EC)m E to the model problem
y ′ = λy ,
y(0) = y0 (6= 0) ,
(84)
where λ < 0, whose solution is, trivially, the decaying exponential y(x) = y0 exp(λx),
x ≥ 0. Our aim is to identify the values of the step size h for which the numerical solution
computed by the P (EC)m E method exhibits a similar exponential decay. The resulting
recursion is
[0]
yn+k +
k−1
X
[m]
¯
α∗j yn+j = h
j=0
[s+1]
yn+k +
k−1
X
k−1
X
[m]
βj∗ yn+j ,
j=0
[m]
¯ k y [s] + h
¯
αj yn+j = hβ
n+k
j=0
k−1
X
[m]
βj yn+j ,
s = 0, . . . , m − 1 ,
j=0
¯ = λh. In order to rewrite this set of equations as a single
for n = 0, 1, 2, . . ., where h
[m] [m]
[m]
difference equation involving yn , yn+1 , . . . yn+k only, we have to eliminate the intermediate
[0]
[m−1]
stages involving yn+k , . . . , yn+k
from the above recursion.
[0]
Let us first take s = 0 and eliminate yn+k form the resulting pair of equations to obtain
[1]
yn+k
+
k−1
X
[m]
αj yn+j
j=0

¯ k h
¯
= hβ
k−1
X
[m]
βj∗ yn+j
j=0
−
k−1
X
j=0

[m]
α∗j yn+j 
¯
+h
k−1
X
[m]
βj yn+j .
j=0
[1]
Now take s = 1 and use the last equation to eliminate yn+k ; this gives,
[2]
yn+k +
k−1
X
j=0


[m]
¯ k hβ
¯ k h
¯
αj yn+j = hβ
¯
+h
k−1
X
j=0
k−1
X
j=0
[m]
βj yn+j −
[m]
βj∗ yn+j −
k−1
X
j=0
Equivalently,
[2]
¯ k
yn+k + 1 + hβ
X
k−1
k−1
X
j=0
[m]

[m]

k−1
X
βj yn+j .
k−1
X
βj yn+j .
α∗j yn+j 
¯
αj yn+j  + h
[m]
j=0
[m]
αj yn+j
j=0

¯ k )2 h
¯
= (hβ
k−1
X
j=0
[m]
βj∗ yn+j −
k−1
X
j=0

[m]
¯ k )h
¯
α∗j yn+j  + (1 + hβ
50
j=0
[m]
By induction,
[m]
¯ k + · · · + (hβ
¯ k )m−1
yn+k + 1 + hβ

¯
¯ k )m h
= (hβ
k−1
X
j=0
[m]
βj∗ yn+j −
k−1
X
j=0
k−1
X
[m]
αj yn+j
j=0

[m]
¯
¯ k + · · · + (hβ
¯ k )m−1 h
α∗j yn+j  + 1 + hβ
[m]
k−1
X
[m]
βj yn+j .
j=0
[m]
For m fixed, this is a kth order difference equation involving yn , . . . , yn+k . In order to
ensure that the solution to this exhibits exponential decay as n → ∞, we have to assume
that all roots to the associated characteristic equation
¯ k )m−1
zk + 1 + ¯
hβk + · · · + (hβ

¯ k )m h
¯
= (hβ
k−1
X
βj∗ z j
j=0
−
k−1
X
j=0
k−1
X
αj z j
j=0

α∗j z j  +
¯ k + · · · + (hβ
¯ k )m−1 h
¯
1 + hβ
k−1
X
βj z j
j=0
have modulus < 1. This can be rewritten in the equivalent form
¯ k + · · · + (hβ
¯ k )m−1 (ρ(z) − hσ(z))
¯
¯ k )m ρ∗ (z) − ¯
Az k + 1 + hβ
+ (hβ
hσ ∗ (z) = 0 ,
where
¯ k )m−1 (hβ
¯ k − αk ) + (hβ
¯ k )m (hβ
¯ ∗ − α∗ ) ,
A =1+ 1+¯
hβk + · · · + (hβ
k
k
Now, since αk = α∗k = 1 and βk∗ = 0, we deduce that A = 0, and therefore the characteristic
equation of the P (EC)m E method is
¯ ≡ ρ(z) − hσ(z)
¯
¯ ∗ (z) − hσ
¯ ∗ (z)) = 0 ,
πP (EC)m E (z; h)
+ Mm (h)(ρ
where
¯ =
Mm (h)
¯ k )m
(hβ
¯ k + · · · + (hβ
¯ k )m−1 ,
1 + hβ
m≥1.
¯ is referred to as the stability polynomial of the predictor-corrector
Here, πP (EC)m E (z; h)
m
method P (EC) E.
By pursuing a similar argument we can also deduce that the characteristic equation of
the predictor corrector method P (EC)m is
¯
Mm (h)
¯ ≡ ρ(z) − hσ(z)
¯
πP (EC)m (z; h)
+ ¯
(ρ∗ (z)σ(z) − ρ(z)σ ∗ (z)) = 0 .
hβk
¯ is referred to as the stability polynomial of the predictor-corrector
Here, πP (EC)m (z; h)
m
method P (EC) .
With the predictor and corrector specified, one can now check using the Schur criterion
or the Routh–Hurwitz criterion, just as in the case of a single multi-step method, whether
¯ and πP (EC)m (z; h)
¯ all lie in the open unit disk |z| < 1 thereby
the roots of πP (EC)m E (z; h)
ensuring the absolute stability of the P (EC)m E and P (EC)m method, respectively.
51
¯ k | < 1, i.e. that 0 < h < 1/|λβk |; then,
Let us suppose, for example, that |hβ
¯ = 0, and consequently,
limm→∞ Mm (h)
¯ = π(z, h)
¯ ,
lim πP (EC)m E (z; h)
¯ = π(z, h)
¯ ,
lim πP (EC)m (z; h)
m→∞
m→∞
¯ = ρ(z) − hσ(z)
¯
where π(z; h)
is the stability polynomial of the corrector. This implies
that in the mode of correcting to convergence the absolute stability properties of the
¯ k | < 1.
predictor-corrector method are those of the corrector alone, provided that |hβ
3.8.2
The accuracy of predictor-corrector methods
Let us suppose that the predictor P has order of accuracy p∗ and the corrector has order of
accuracy p. The question we would like to investigate here is: What is the overall accuracy
of the predictor-corrector method?
Let us consider the P (EC)m E method applied to the model problem (84) with m ≥ 1.
We have that
¯
¯
¯
¯
¯
¯ = ρ(eh ) − ¯
¯ ∗ (eh ) − hσ
¯ ∗ (eh ))
hσ(eh ) + Mm (h)(ρ
πP (EC)m E (eh ; h)
¯
¯ p∗ +1 )
¯ p+1 ) + Mm (h)O(
h
= O(h
¯ p+1 + h
¯ p∗ +m+1 )
= O(h

¯ p+1 + h
¯ p+2 )

if p∗ ≥ p
 O(h
¯ p+1 )
=
O(h
if p∗ = p − q, 0 < q ≤ p and m ≥ q

 O(h
¯ p+1 + h
¯ p−q+m+1 ) if p∗ = p − q, 0 < q ≤ p and m ≤ q − 1 .
P (EC)m E
Consequently, denoting by Tn
have that
m
TnP (EC) E
the truncation error of the method P (EC)m E, we

¯p)

 O(h
if p∗ ≥ p
=
if p∗ = p − q, 0 < q ≤ p and m ≥ q

 O(h
¯ p−q+m ) if p∗ = p − q, 0 < q ≤ p and m ≤ q − 1 .
¯p)
O(h
This implies that from the point of view of overall accuracy there is no particular advantage
in using a predictor of order p∗ ≥ p. Indeed, as long as p∗ + m ≥ p, the predictor-corrector
method P (EC)m E will have order of accuracy p.
Similar statements can be made about P (EC)m type predictor-corrector methods with
m ≥ 1. On writing
¯ ∗ (z))σ(z) − σ ∗ (z)(ρ(z) − hσ(z))
¯
ρ∗ (z)σ(z) − σ ∗ (z)ρ(z) = (ρ∗ (z) − hσ
,
we deduce that
¯ ¯
¯ p+1 + h
¯ p∗ +m + h
¯ p+m ) .
πP (EC)m (eh ; h)
= O(h
Consequently, as long as p∗ + m ≥ p + 1 the predictor-corrector method P (EC)m will have
order of accuracy p.
52
4
Stiff problems
Let us consider an initial value problem for a system of m ordinary differential equations
of the form:
y′ = f (x, y) ,
y(a) = y0 ,
(85)
where y = (y1 , . . . , ym )T . A linear k-step method for the numerical solution of (85) has
the form
k
X
αj yn+j = h
j=0
k
X
βj fn+j ,
(86)
j=0
where fn+j = f (xn+j , yn+j ). Let us suppose, for simplicity, that f (x, y) = Ay + b where
A is a constant matrix of size m × m and b is a constant (column) vector of size m; then
(86) becomes
k
X
(αj I − hβj A)yn+j = hσ(1)b ,
(87)
j=0
where σ(1) = kj=0 βj (6= 0) and I is the m × m identity
eigenvalues λi , i = 1, . . . , m, of the matrix A are distinct.
matrix H such that

λ1 0 · · ·
 0 λ

2 ···
HAH −1 = Λ = 
 ··· ··· ···
0
0 ···
P
matrix. Let us suppose that the
Then, there exists a nonsingular
0
0
···
λm



 .

Let us define z = Hy and c = hσ(1)Hb. Thus, (87) can be rewritten as
k
X
(αj I − hβj Λ)zn+j = c ,
(88)
j=0
or, in component-wise form,
k
X
(αj − hβj λi )zn+j,i = ci ,
j=0
where zn+j,i and ci , i = 1, . . . , m, are the components of zn+j and c respectively. Each
of these m equations is completely decoupled from the other m − 1 equations. Thus we
are now in the framework of Section 3 where we considered linear multistep methods for a
single differential equation. However, there is a new feature here: given that the numbers
λi , i = 1, . . . , m, are eigenvalues of the matrix A, they need not be real numbers. As a
¯ = hλ, where λ is any of the m eigenvalues, can be a complex
consequence the parameter h
number. This leads to the following modification of the definition of absolute stability.
Definition 10 A linear k-step method is said to be absolutely stable in an open set RA
¯ ∈ RA all roots rs , s = 1, . . . , k, of the stability polynomial
of the complex plane, if for all h
¯
π(r, h) associated with the method, and defined by (78), satisfy |rs | < 1. The set RA is
called the region of absolute stability of the method.
Clearly, the interval of absolute stability of a linear multistep method is a subset of its
region of absolute stability.
53
Exercise 6
a) Find the region of absolute stability of Euler’s explicit method when applied to the
ordinary differential equation y ′ = λy, y(x0 ) = y0 .
b) Suppose that Euler’s explicit method is applied to the second-order differential equation
y(0) = 1 , y ′ (0) = λ − 2 ,
y ′′ + (λ + 1)y ′ + λy = 0 ,
rewritten as a first-order system in the vector (u, v)T , with u = y and v = y ′ . Suppose
that λ >> 1. What choice of the step size h will guarantee absolute stability in the
sense of Definition 10?
¯ = ρ(z) −
Solution: a) For Euler’s explicit method ρ(z) = z − 1 and σ(z) = 1, so that π(z; h)
¯
¯ = z − (1 + h).
¯ This has the root r = (1 + ¯h). Hence the region of absolute
hσ(z)
= (z − 1) − h
¯ ∈ C : |1 + h|
¯ < 1} which is an open unit circle centred at −1.
stability is RA = {h
b) Now writing u = y and v = y ′ and y = (u, v)T , the initial value problem for the given
second-order differential equation can be recast as
y′ = Ay ,
where
A=
0
1
−λ −(λ + 1)
y(0) = y0 ,
and y0 =
1
λ−2
.
The eigenvalues of A are the roots of the characteristic polynomial of A,
det(A − zI) = z 2 + (λ + 1)z + λ .
Hence, r1 = −1, r2 = −λ; we deduce that the method is absolutely stable provided that h < 2/λ.
It is an easy matter to show that
u(x) = 2e−x − e−λx ,
v(x) = −2e−x + λe−λx .
The graphs of the functions u and v are shown in the figure below for λ = 45. Note that v exhibits
a rapidly varying behaviour (fast time scale) near x = 0 while u is slowly varying for x > 0 and
v is slowly varying for x > 1/45. Despite the fact that over most of the interval [0, 1] both u and
v are slowly varying when λ = 45, we are forced to use a step size of h < 2/45 in order to ensure
that the method is absolutely stable. ⋄
In the example the v component of the solution exhibited two vastly different time
scales; in addition, the fast time scale (which occurs between x = 0 and x ≈ 1/λ) has
negligible effect on the solution so its accurate resolution does not appear to be important
for obtaining an overall accurate solution. Nevertheless, in order to ensure the stability
of the numerical method under consideration, the mesh size h is forced to be exceedingly
small, h < 2/λ, smaller than an accurate approximation of the solution for x ≫ 1/λ would
necessitate. Systems of differential equations which exhibit this behaviour are generally
referred to as stiff systems. We refrain from formulating a rigorous definition of stiffness.
4.1
Stability of numerical methods for stiff systems
In order to motivate the various definitions of stability which occur in this section, we begin
with a simple example. Consider Euler’s implicit method for the initial value problem
y ′ = λy ,
y(0) = y0 ,
54
40
30
20
10
0
0.2
0.4
x
0.6
0.8
1
Figure 2: The functions u and v plotted against x for x ∈ [0, 1].
¯ = ρ(z) −
where λ is a complex number. The stability polynomial of the method is π(z, h)
¯
¯
hσ(z) where h = hλ, ρ(z) = z − 1 and σ(z) = z. Since the only root of the stability
polynomial is z = 1/(1 − ¯
h), we deduce that the method has the region of stability
¯ : |1 − h|
¯ > 1} .
R = {h
In particular R includes the whole of the left-hand complex half plane. The next definition
is due to Dahlquist (1963).
Definition 11 A linear multistep method is said to be A-stable if its region of absolute
stability, RA contains the whole of the left-hand complex half-plane ℜ(hλ) < 0.
As the next theorem by Dahlquist (1963) shows, Definition 11 is far too restrictive.
Theorem 16
(i) No explicit linear multistep method is A-stable.
(ii) The order of an A-stable implicit linear multistep method cannot exceed 2.
(iii) The second-order A-stable linear multistep method with smallest error constant is
the trapezium rule.
Thus we adopt the following definition due to Widlund (1967).
55
Definition 12 A linear multistep method is said to be A(α)-stable, α ∈ (0, π/2), if its
region of absolute stability RA contains the infinite wedge
¯ | π − α < arg(h)
¯ < π + α} .
Wα = {h
A linear multistep method is said to be A(0)-stable if it is A(α)-stable for some α ∈
(0, π/2). A linear multistep method is A0 stable if RA includes the negative real axis in
the complex plane.
Let us note in connection with this definition that if ℜλ < 0 for a given λ then
¯ = hλ either lies inside the wedge Wα or outside Wα for all positive h. Consequently,
h
if all eigenvalues λ of the matrix A happen to lie in some wedge Wβ then an A(β)-stable
method can be used for the numerical solution of the initial value problem (85) without
any restrictions on the step size h. In particular, if all eigenvalues of A are real and
nonnegative, then an A(0) stable method can be used. The next theorem (stated here
without proof) can be regarded the analogue of Theorem 16 for the case of A(α) and A(0)
stability.
Theorem 17
(i) No explicit linear multistep method is A(0)-stable.
(ii) The only A(0)-stable linear k-step method whose order exceeds k is the trapezium
rule.
(iii) For each α ∈ [0, π/2) there exist A(α)-stable linear k-step methods of order p for
which k = p = 3 and k = p = 4.
A different way of loosening the concept of A-stability was proposed by Gear (1969).
The motivation behind it is the fact that for a typical stiff problem the eigenvalues of the
¯ = −a, a > 0, in
matrix A which produce the fast transients all lie to the left of a line ℜh
the complex plane, while those that are responsible for the slow transients are clustered
around zero.
Definition 13 A linear multistep method is said to be stiffly stable if there exist positive
real numbers a and c such that RA ⊃ R1 ∪ R2 where
¯ ∈ C : ℜh
¯ < −a}
R1 = {h
and
¯ ∈ C : −a ≤ ℜh
¯ < 0, −c ≤ ℑh
ˆ ≤ c} .
R2 = {h
It is clear that stiff stability implies A(α)-stability with α = arctan(c/a). More generally, we have the following chain of implications:
A-stability ⇒ stiff-stability ⇒ A(α)-stability ⇒ A(0)-stability ⇒ A0 -stability .
In the next two sections we shall consider linear multistep methods which are particularly well suited for the numerical solution of stiff systems of ordinary differential
equations.
56
k
α6
α5
α4
α3
α2
α1
α0
βk
p
Cp+1
amin
αmax
1
−1
1
1
− 12
0
90o
1
− 43
1
3
2
3
2
− 29
0
90o
1
− 18
11
9
11
2
− 11
6
11
3
3
− 22
0.1
88o
1
− 48
25
36
25
− 16
25
3
25
12
25
4
12
− 125
0.7
73o
1
− 300
137
300
137
− 200
137
75
137
12
− 137
60
137
5
10
− 137
2.4
52o
− 360
147
450
147
− 400
147
225
147
72
− 147
10
147
60
147
6
20
− 343
6.1
19o
1
2
3
4
5
6
1
Table 3: Coefficients, order, error constant and stability parameters for backward differentiation methods
4.2
Backward differentiation methods for stiff systems
¯ = ρ(z) − hσ(z).
¯
Consider a linear multistep method with stability polynomial π(z, h)
If
¯
¯
the method is A(α)-stable or stiffly stable, the roots r(h) of π(·, h) lie in the closed unit
¯ is real and h
¯ → −∞. Hence,
disk when h
¯
ρ(r(h))
¯ = σ( lim r(h))
¯ ;
= lim σ(r(h))
¯
¯
¯
¯
h
h→−∞
h→−∞
h→−∞
0 = lim
¯ approach those of σ(·). Thus it is natural to choose σ(·)
in other words, the roots of π(·, h)
in such a way that its roots lie within the unit disk. Indeed, a particularly simple choice
would be to take σ(z) = βk z k ; the resulting class of, so-called, backward differentiation
methods has the general form:
k
X
αj yn+j = hβk fn+k
j=0
where the coefficients αj and βk are given in Table 3 which also displays the value of a in
the definition of stiff stability and the angle α from the definition of A(α) stability, the
order p of the method and the corresponding error constant Cp+1 for p = 1, . . . , 6. For
p ≥ 7 backward differentiation methods of order p of the kind considered here are not
zero-stable and are therefore irrelevant from the practical point of view.
57
4.3
Gear’s method
Since backward differentiation methods are implicit, they have to be used in conjunction
with a predictor. Instead of iterating the corrector to convergence via a fixed point iteration, Newton’s method can be used to accelerate the iterative convergence of the corrector.
Rewriting the resulting predictor-corrector multi-step pair as a one step method gives rise
to Gear’s method which allows the local alteration of the order of the method as well
as of the mesh size. We elaborate on this below.
As we have seen in Section 4.1, in the numerical solution of stiff systems of ordinary
differential equations, the stability considerations highlighted in parts (i) of Theorems 16
and 17 necessitate the use of implicit methods. Indeed, if a predictor-corrector method is
used with a backward differentiation formula as corrector, a system of nonlinear equations
of the form
yn+k − hβk f (xn+k , yn+k ) = gn+k
will have to be solved at each step, where
gn+k = −
k−1
X
αj yn+j
j=0
is a term that involves information which has already been computed at previous steps
and can be considered known. If this equation is solved by a fixed-point iteration, the
Contraction Mapping Theorem will require that
Lh|βk | < 1
(89)
in order to ensure convergence of the iteration; here L is the Lipschitz constant of the
function f (x, ·). In fact, since the function f (x, ·) is assumed to be continuously differentiable,
∂f
L = max (x, y) .
(x,y)∈R ∂x
For a stiff system L is typically very large, thus the restriction on the steplength h expressed
by (89) is just as severe as the condition on h that one encounters when using an explicit
method with a bounded region of absolute stability. In order to overcome this difficulty,
Gear proposed to use Newton’s method:
[s+1]
yn+k
=
[s]
yn+k
−1 h
∂f
[s]
− I − hβk (xn+k , yn+k )
∂y
[s]
[s]
i
yn+k − hβk f (xn+k , yn+k ) − gn+k , (90)
[0]
for s = 0, 1, . . . , with a suitable initial guess yn+k . Even when applied to a stiff system,
convergence of the Newton iteration (90) can be attained without further restrictions on
[0]
the mesh size h provided that we can supply a sufficiently accurate initial guess yn+k (by
using an appropriately accurate predictor, for example).
On the other hand, the use of Newton’s method in this context has the disadvantage
[s]
∂f
that the Jacobi matrix ∂f /∂y has to be reevaluated and the matrix I − hβk ∂y
(xn+k , yn+k )
inverted at each step of the iteration and at each mesh point xn+k .
[s]
∂f
One aspect of Gear’s method is that the matrix I − hβk ∂y
(xn+k , yn+k ) involved in the
Newton iteration described above is only calculated occasionally (i.e. at the start of the
58
iteration, for s = 0, and thereafter only if the Newton iteration exhibits slow convergence);
the inversion of this matrix is performed by an LU decomposition.
A further aspect of Gear’s method is a strategy for varying the order of the backward
differentiation formula and the step size according to the intermediate results in the computation. This is achieved by rewriting the multistep predictor-corrector pair as a one-step
method (in the so-called Nordsieck form). For further details, we refer to Chapter III.6 in
the book of Hairer, Norsett and Wanner.
5
Nonlinear Stability
All notions of stability which were considered so far in these notes rest on the assumption
that f (x, y) = λy or f (x, y) = Ay + b. The purpose of this section is to develop an
appropriate theoretical framework which is directly applicable to the stability analysis of
numerical methods for nonlinear ODEs.
Consider the linear system of differential equations y′ = Ay, y(x0 ) = y0 , where all
eigenvalues of the m × m matrix A have negative real part. Then ky(x)k decreases as x
increases; also, neighbouring solution curves get closer as x increases: if u(x) and v(x) are
two solutions to y′ = Ay subject to u(x0 ) = u0 and v(x0 ) = v0 , respectively, then
ku(x2 ) − v(x2 )k ≤ e(x2 −x1 ) maxi ℜλi (A) ku(x1 ) − v(x1 )k ,
x2 ≥ x1 ≥ x0 ,
(91)
where k · k denotes the Euclidean norm on Rm . Clearly,
0 < e(x2 −x1 ) maxi ℜλi (A) ≤ 1
for x2 ≥ x1 ≥ x0 and therefore,
ku(x2 ) − v(x2 )k ≤ ku(x1 ) − v(x1 )k ,
x2 ≥ x1 ≥ x0 .
It is the property (91) that has a natural extension to nonlinear differential equations and
leads to the following definition.
Definition 14 Suppose that u and v are two solutions of the differential equation y′ =
f (x, y) subject to respective initial conditions u(x0 ) = u0 , v(x0 ) = v0 . If
ku(x2 ) − v(x1 )k ≤ ku(x1 ) − v(x1 )k
for all real numbers x2 and x1 such that x2 ≥ x1 ≥ x0 , where k · k denotes the Euclidean
norm on Cm , then the solutions u and v are said to be contractive in the norm k · k.
To see what assumptions of f we must make to ensure that two solutions u and v
to the differential equations y′ = f (x, y), with respective initial conditions u(x0 ) = u0 ,
v(x0 ) = v0 , are contractive, let h·, ·i denote the inner product of Rm and consider the
inner product of u′ − v′ = f (x, u) − f (x, v) with u − v. This yields,
1 d
ku(x) − v(x)k2 = hf (x, u(x)) − f (x, v(x)), u(x) − v(x)i .
2 dx
Thus, if
hf (x, u(x)) − f (x, v(x)), u(x) − v(x)i ≤ 0
59
(92)
for all x ≥ x0 then
1 d
ku(x) − v(x)k2 ≤ 0
2 dx
for all x ≥ x0 , and therefore the solutions u and v corresponding to initial conditions u0
and v0 , respectively, are contractive in the norm k · k. The inequality (92) motivates the
following definition.
Definition 15 The system of differential equations y′ = f (x, y) is said to be dissipative
in the interval [x0 , ∞) if
hf (x, u) − f (x, v), u − vi ≤ 0
(93)
for all x ≥ x0 and all u and v in Rm .
Thus we have proved that if the system of differential equations is dissipative then any
solutions u and v corresponding to respective initial values u0 and v0 are contractive. A
slight generalisation of (92) results in the following definition.
Definition 16 The function f (x, ·) is said to satisfy a one-sided Lipschitz condition
on the interval [x0 , ∞) if there exists a function ν(x) such that
hf (x, u) − f (x, v), u − vi ≤ ν(x)ku − vk2
(94)
for all x ∈ [x0 , ∞).
In particular, if f (x, ·) satisfies a one-sided Lipschitz condition on [x0 , ∞) and ν(x) ≤ 0 for
all x ∈ [x0 , ∞), then the differential equation y′ = f (x, y) is dissipative on this interval,
and therefore any pair of solutions u and v to this equations, corresponding to respective
initial values u0 and v0 are also contractive.
Now we shall consider numerical methods for the solution of an initial value problem
for a dissipative differential equation y′ = f (x, y) and develop conditions under which numerical solutions are also contractive in a suitable norm. In order to keep the presentation
simple we shall suppose that f is independent of x and, instead of a linear k-step method,
k
X
αj yn+j = h
j=0
k
X
βj f (yn+j )
(95)
j=0
for the numerical solution of y′ = f (y), y(x0 ) = y0 , we shall consider its one-leg twin
k
X
j=0

αj yn+j = hf 
k
X
j=0

βj yn+j  .
(96)
For example, the one-leg twin of the trapezium rule method
1
yn+1 − yn = h [f (yn+1 ) + f (yn )]
2
is the implicit midpoint rule method
yn+1 − yn = hf
1
(yn+1 + yn )
2
60
.
Let us recall the notation from Section 3.1 to simplify writing: putting yn+1 = Eyn , we
can write the linear k-step method (95) as
ρ(E)yn = hσ(E)f (yn ) ,
where ρ(z) = kj=0 αj z j and σ(z) = kj=0 βj z j are the first and second characteristic
polynomial of the method; the associated one-leg twin (96) is then
P
P
ρ(E)yn = hf (σ(E)yn ) .
There is a close relationship between the linear multistep method (95) and its one-leg
twin (96). Let zn = σ(E)yn ; then
ρ(E)zn = ρ(E)σ(E)yn = σ(E)ρ(E)yn = hσ(E)f (σ(E)yn ) = hσ(E)f (zn ) .
In other words, if {yn }n≥0 is a solution to (96) then {zn }n≥0 , with zn = σ(E)yn , is the
solution of the linear multistep method (95) whose one-leg twin (96) is. This connection
allows results for the one-leg twin to be translated into results for the linear multistep
method.
Now we shall state a definition of nonlinear stability due to Dahlquist (1975). Before
we do so, we introduce the concept of G-norm. Consider a vector
Zn = (zTn+k−1 , . . . , zTn )T
in Rmk where zn+j ∈ Rm , j = 0, 1, . . . , k − 1. Given that G = (gij ) is a k × k symmetric
positive definite matrix, the G-norm k · kG is defined by
kZn kG

1/2
k X
k
X
=
gij hzn+k−i , zn+k−j i
.
i=1 j=1
Definition 17 The k-step method (96) is said to be G-stable if there exists a symmetric
positive definite matrix G such that
kZn+1 k2G − kZn k2G ≤ 2hρ(E)zn , σ(E)zn i/σ 2 (1) .
Let {un } and {vn } be two solutions of the differential equation y′ = f (y) given by
(96) with different starting values, and suppose that (96) is G-stable. Define the vectors
Un and Vn in Rmk by
Un = (uTn+k−1 , . . . , uTn )T ,
T
Vn = (vn+k−1
, . . . , vnT )T .
Since we are assuming that the differential equation y′ = f (y) is dissipative, it follows
that
kUn+1 − Vn+1 k2G − kUn − Vn k2G ≤ 2hρ(E)(un − vn ), σ(E)(un − vn )i/σ 2 (1)
≤ 2hhf (σ(E)un ) − hf (σ(E)vn ), σ(E)(un − vn )i/σ 2 (1)
=
2h
hf (σ(E)un ) − hf (σ(E)vn ), σ(E)un − σ(E)vn )i ≤ 0 .
σ(1)2
61
Hence,
kUn+1 − Vn+1 k2G ≤ kUn − Vn k2G
for n ≥ 0 .
Thus we deduce that a G-stable approximation of a dissipative ordinary differential equation y′ = f (y) is contractive in the G-norm; this is a discrete counterpart of the property
we established at the beginning of this section that analytical solutions to a dissipative
equation of this kind are contractive in the Euclidean norm. For further developments of
these ideas, we refer to the books of J.D. Lambert, and A.M. Stuart and A.R. Humphries.
6
Boundary value problems
In many applications a system of m simultaneous first-order ordinary differential equations
in m unknowns y1 (x), y2 (x), . . . , ym (x) has to be solved. If each of these variables satisfies
a given condition at the same value a of x then we have an initial value problem for a
system of first-order ordinary differential equations. If the yi , i = 1, . . . , m, satisfy given
conditions at different values a, b, c, . . . of the independent variable x then we have a
multi-point boundary value problem. In particular, if conditions on the yi , i = 1, . . . , m,
are imposed at two different values a and b then we have a two-point boundary value
problem.
Example 7 Here is an example of a multipoint (in this case, three-point) boundary value
problem:
y ′′′ − y ′′ + y ′ − y = 0 ,
y(0) = 1 , y(1) = e , y ′ (2) = e2 .
The exact solution is y(x) = ex .
Example 8 This is an example of a two-point boundary value problem:
y ′′ − 2y 3 = 0 ,
y(1) = 1 ,
y ′ (2) + [y(2)]2 = 0 .
The exact solution is y(x) = 1/x.
In this section we shall consider three classes of methods for the numerical solution of
two-point boundary value problems: shooting methods, matrix methods and collocation
methods.
6.1
Shooting methods
Let us consider the two-point boundary value problem
y ′′ = f (x, y, y ′ ) ,
y(a) = A ,
y(b) = B ,
(97)
with a < b and x ∈ [a, b]. We shall suppose that (97) has a unique solution. The motivation
behind shooting methods is to convert the two-point boundary value problem into solving
a sequence of initial value problems whose solutions converge to that of the boundary
value problem, so that one can use existing software developed for the numerical solution
of initial value problems: observe that an attempt to solve the boundary value problem
(97) directly will lead to a coupled system of nonlinear equations whose solution may be
a hard problem.
62
Let us make an initial guess s for y ′ (a) and denote by y(x; s) the solution of the initial
value problem
y ′′ = f (x, y, y ′ ) ,
y(a) = A , y ′ (a) = s .
(98)
Introducing the notation u(x; s) = y(x; s), v(x; s) =
system of first-order ordinary differential equations:
∂
∂x y(x; s),
∂
u(x; s) = v(x; s) ,
∂x
we can rewrite (98) as a
u(a; s) = A ,
(99)
∂
v(x; s) = f (x, u(x; s), v(x; s)) ,
∂x
v(a; s) = s .
The solution u(x; s) of the initial value problem (99) will coincide with with the solution
y(x) of the boundary value problem (97) provided that that we can find a value of s such
that
φ(s) ≡ u(b; s) − B = 0 .
(100)
The essence of the shooting method for the numerical solution of the boundary value
problem (97) is to find a root to the equation (100). Any standard root-finding technique
can be used; here we shall consider two: bisection of the interval which is known to contain
the root and Newton’s method.
6.1.1
The method of bisection
Let us suppose that that two numbers s1 and s2 are known such that
φ(s1 ) < 0
and
φ(s2 ) > 0 .
We assume, for the sake of definiteness, that s1 < s2 . Given that the solution of the initial
value problem (99) depends continuously on the initial data, there must exist at least one
value of s in the interval (s1 , s2 ) such that φ(s) = 0. Thus the interval [s1 , s2 ] contains a
root of the equation (100).
The root of (100) can now be calculated approximately using the method of bisection.
We take the midpoint s3 of the interval [s1 , s2 ], compute u(b, s3 ) and consider whether
φ(s3 ) = u(b; s3 ) − B is positive or negative. If φ(s3 ) > 0 then it is the interval [s1 , s3 ]
that contains a root of φ, whereas if φ(s3 ) < 0 then the interval in question is [s3 , s2 ]. By
repeating this process, one can construct a sequence of numbers {sn }∞
n=1 converging the
s. In practice the process of bisection is terminated after a finite number of steps when
the length of the interval containing s has become sufficiently small.
6.1.2
The Newton–Raphson method
An alternative to the method of bisection is to compute a sequence {sn }∞
n=1 generated by
the Newton–Raphson method:
sn+1 = sn − φ(sn )/φ′ (sn ) ,
(101)
with the starting value s0 chosen arbitrarily in a sufficiently small interval surrounding the
root. For example, a suitable s0 may be found by performing a few steps of the method of
63
bisection. If s0 is a sufficiently good approximation to the required root of (100) the theory
of the Newton–Raphson method ensures that, in general, we have quadratic convergence
of the sequence {sn }∞
n=0 to the root s.
From the point of view of implementing (101) the first question that we need to clarify
is how one can compute compute φ′ (sn ). To do so, we introduce the new dependent
variables
∂u(x; s)
∂v(x; s)
,
η(x; s) =
ξ(x; s) =
∂s
∂s
and differentiate the initial value problem (99) with respect to s to obtain a second initial
value problem:
∂ξ(x; s)
∂x
= η(x; s) ,
ξ(a; s) = 0 ,
(102)
∂η(x; s)
∂x
= p(x; s)ξ(x; s) + q(x; s)η(x; s) ,
η(a; s) = 1 ,
where
p(x; s) =
∂f (x, u(x; s), v(x; s))
,
∂u
(103)
q(x; s) =
∂f (x, u(x; s), v(x; s))
.
∂v
If we assign the value sn to s, n ≥ 0, then the initial value problem (99), (102) can be solved
by a predictor-corrector method or a Runge–Kutta method on the interval [a, b]. Thus we
obtain u(b; sn ) or, more precisely, an approximation to u(b; sn ) from which we can calculate
φ(sn ) = u(b; sn ) − B; in addition, we obtain an approximation to ξ(b; sn ) = φ′ (sn ). Having
calculated φ(sn ) and φ′ (sn ), we obtain the next Newton–Raphson iterate sn+1 from (101).
The process is then repeated until the iterates sn settle to a fixed number of digits.
Two remarks are in order:
1) According to (103), the initial value problems (99) and (102) are coupled and therefore they must be solved simultaneously over the interval [a, b], with s set to sn ,
n = 0, 1, 2, . . .;
2) The coupled initial value problem (99), (102) may be very sensitive to perturbations
of the initial guess s0 ; a bad initial guess of s0 may result in a sequence of Newton–
Raphson iterates {sn }∞
n=0 which does not converge to the root s.
The latter difficulty may be overcome by using the multiple shooting method
which we describe below. First, however, we show how the simple shooting method may
be extended to the nonlinear two-point boundary value problem
y′ = f (x, y) ,
a<x<b,
g(y(a), y(b)) = 0 ,
(104)
where y, f and g are m-component vector functions of their respective arguments. In the
simple shooting method the boundary value problem (104) is transformed into the initial
value problem
u′ (x; s) = f (x, u(x; s)) ,
a<x<b,
64
u(a; s) = s ,
(105)
whose solution is required to satisfy the condition
G(s) ≡ g(s, u(b; s)) = 0
(106)
for an unknown value of s. Thus the problem has been transformed into one of finding a
solution to the system of nonlinear equations G(s) = 0. In order to evaluate the function
G for a specific value of s a numerical method for the solution of the initial value problem
(105) has to be employed on the interval [a, b] to compute (an approximation to) u(b; s).
In fact, as we have already seen earlier, a root-finding procedure such as Newton’s method
will require the computation of the Jacobi matrix
J(s) =
∂G
.
∂s
This in turn requires the solution of a coupled initial value problem for m2 linear ordinary
differential equations to obtain ∂u/∂s for a ≤ x ≤ b.
The shooting method is said to converge if the root-finding algorithm results in a
sequence {sn }∞
n=0 which converges to a root s of G; then s = y(a) where y(x) is the
desired solution of the boundary value problem (104). The convergence of this sequence
and therefore the success of the shooting method may be hampered by two effects:
1) A well-conditioned boundary value problem of the form (104) may easily lead to an
ill-posed initial value problem (104);
2) A bounded solution to the initial value problem (104) may exist only for s in a small
neighbourhood of y(s) (which, of course, in unknown).
The idea behind the multiple shooting method is to divide the interval [a, b] into
smaller subintervals
(107)
a = x0 < x1 < · · · < xK−1 < xK = b ;
the problem then is to find a vector ST = (sT0 , . . . , sTK−1 ) such that the solutions uk (x; sk )
of the initial value problems
u′k (x; sk ) = f (x, uk (x; sk )) ,
uk (xk ; sk ) = sk ,
xk < x ≤ xk+1 ,
(108)
k = 0, . . . , K − 1 ,
satisfy the conditions
uk (xk+1 ; sk ) − sk+1 = 0 ,
k = 0, . . . , K − 2 ,
g(s0 , uK−1 (b; sK−1 )) = 0 .
(109)
The equations (109) can be written in the compact form G(S) = 0. A clear advantage
of the multiple shooting method over simple shooting is that the growth of the solutions
to the initial value problems (108) and the related linear initial value problems for the
∂uk (x; sk )/∂sk , k = 0, . . . , K−1, can be approximated accurately by selecting a sufficiently
fine subdivision (107) of the interval [a, b].
Indeed, it is possible to prove that, under reasonable assumptions, the multiple shooting method leads to an exponential increase of the size of the domain of initial values
for which the first iteration of the root-finding procedure is defined. This is consistent
with the practical observation that multiple shooting is less sensitive to the choice of the
starting values than simple shooting.
65
6.2
Matrix methods
In this section, rather than attempting to convert the boundary value problem to an
initial value problem, we approximate it directly by using a finite difference method. This
results in a system of equations for the unknown values of the numerical solution at the
mesh points. We begin by considering a linear boundary value problem. In this case
the calculation of the numerical solution amounts to solving a system of linear equations
with a sparse matrix. We then consider a nonlinear boundary value problem; upon the
application of Newton’s method, this again involves, in each Newton iteration, the solution
of a system of linear equations with a sparse matrix. Methods of this kind are usually
referred to in the ODE literature as matrix methods.
6.2.1
Linear boundary value problem
Let us consider the two-point boundary value problem
y ′′ + p(x)y ′ + q(x)y = f (x) ,
x ∈ (a, b) ,
′
a0 y(a) + b0 y (a) = c0 ,
(110)
′
a1 y(b) + b1 y (b) = c1 .
We discretise (110) using a finite difference method on the uniform mesh
{xi | xi = a + ih , i = 0, . . . , N }
of step size h = (b − a)/N , N ≥ 2. The essence of the method is to approximate the
derivatives in the differential equation and the boundary conditions by divided differences.
Assuming that y is sufficiently smooth, it is a simple matter to show using Taylor series
expansions that
y ′′ (xi ) =
y ′ (xi ) =
y ′ (x0 ) =
y ′ (xN ) =
y(xi+1 ) − 2y(xi ) + y(xi−1 )
+ O(h2 ) ,
h2
y(xi+1 ) − y(xi−1 )
+ O(h2 ) ,
2h
−3y(x0 ) + 4y(x1 ) − y(x2 )
+ O(h2 ) ,
2h
y(xN −2 ) − 4y(xN −1 ) + 3y(xN )
+ O(h2 ) .
2h
(111)
(112)
(113)
(114)
Now, using (111–114) we can construct a finite difference method for the numerical solution
of (110); we denote by yi the numerical approximation to y(xi ) for i = 0, . . . , N :
yi+1 − 2yi + yi−1
yi+1 − yi−1
+ p(xi )
+ q(xi )yi = f (xi ) ,
h2
2h
−3y0 + 4y1 − y2
a0 y0 + b0
= c0 ,
2h
yN −2 − 4yN −1 + 3yN
a1 yN + b1
= c1 .
2h
i = 1, . . . , N − 1 ,
After rearranging these, we obtain
1
p(xi )
−
yi−1 −
h2
2h
2
− q(xi ) yi +
h2
1
p(xi )
+
yi+1 = f (xi ),
h2
2h
66
1≤i≤N −1,
3b0
4b0
b0
a0 −
y0 +
y1 −
y2 = c0 ,
2h
2h
2h
b1
4b1
3b1
yN −2 −
yN −1 + a1 +
yN −2 = c1 .
2h
2h
2h
This is a system of linear equations of the form
A0 y0 + C0 y1 + B0 y2 = F0 ,
Ai yi−1 + Ci yi + Bi yi+1 = Fi ,
AN yN −2 + CN yN −1 + BN yN
= FN .
i = 1, . . . , N − 1 ,
The matrix of the system is





M =



A0
A1
0
···
0
0
C0
C1
A2
···
0
0
B0 0
0
···
0
B1 0
0
···
0
C2 B2
0
···
0
··· ···
···
···
···
0
0 AN −1 CN −1 BN −1
0
0
AN
CN
BN
Let us consider the following cases:





 .



1) If B0 = 0 and AN = 0 then M is a tridiagonal matrix.
2) If B0 6= 0 and B1 = 0 (and/or AN 6= 0 and AN −1 = 0) then we can interchange
the first two rows of the matrix (and/or the last two rows) and, again, we obtain a
tridiagonal matrix.
3) If B0 6= 0 and B1 6= 0 (and/or AN 6= 0 and AN −1 6= 0) then we can eliminate B0
from the first row (and/or AN from the last row) to obtain a tridiagonal matrix.
In any case the matrix M can be transformed into a tridiagonal matrix. Thus, from now
on, without any restrictions on generality, we shall suppose that M is tridiagonal. To
summarise the situation, we wish to solve the system of linear equations
My = F
where M is a tridiaginal matrix,
y = (y0 , y1 , . . . , yN )T ,
F = (F0 , F1 , . . . , FN )T .
The algorithm we present below for the solution of this linear system is usually referred
to as the Thomas algorithm. It is a special case of Gaussian elimination or LU
decomposition; in fact, it is this latter interpretation that we adopt here.
We wish to express M as a product LU of a lower-triangular matrix





L=



1
0
0
0
0
l1
1
0
0
0
0 l2
1
0
0
··· ··· ··· ··· ···
0
0
0
0 lN −1
0
0
0
0
0
67
··· 0
··· 0
··· 0
··· ···
1
0
lN 1









and an upper-triangular matrix





U =



u0 v0 0
0
0
···
0
0 u1 v1 0
0
···
0
0
0 u2 v2 0
···
0
··· ··· ··· ··· ··· ···
···
0
0
0
0
0 uN −1 vN −1
0
0
0
0
0
0
uN





 .



On multiplying L and U and equating the resulting matrix with M , we find that
u0 = C0 ,
v0 = B0 ,
Hence
vi = Bi ,
i = 0, . . . , N ,
u0 = C0 ,

li ui−1 = Ai 

li vi−1 + ui = Ci

vi = Bi 
i = 1, . . . , N .
ui = Ci − (Ai Bi−1 )/ui−1
li = Ai /ui−1
)
i = 1, . . . , N .
Given that the entries of the matrix M are known, the components of L and U can be
computed from these formulae, and the set of linear equations M y = F can be written in
the following equivalent form:
(
Lz = F
.
Uy = z
Thus, instead of solving M y = F directly, we solve two linear systems in succession, each
with a triangular matrix: Lz = F is solved for z, followed by solving U y = z for y.
Writing this out in detail yields
(
and
(
z1 = F1 ,
zi = Fi − li−1 zi−1 ,
yN = zN /uN ,
yi = (zi − vi yi+1 )/ui ,
i = 1, . . . , N ,
i = N − 1, . . . , 0 .
Expressing in these formulae the values of ui and vi in terms of Ai , Bi , Ci , we find that
yi = αi+1 yi+1 + βi+1 ,
yN
= βN +1 ,
i = N − 1, . . . , 0 ,
where
Bi
,
Ci + αi Ai
Fi − βi Ai
=
,
Ci + αi Ai
αi+1 = −
i = 1, 2, . . . , N − 1 ,
α1 = −
βi+1
i = 1, 2, . . . , N ,
β1 =
B0
,
C0
F0
.
C0
The last set of formulae are usually referred to as the Thomas algorithm.
68
6.2.2
Nonlinear boundary value problem
Now consider a nonlinear second-order differential equation subject to linear boundary
conditions:
y ′′ = f (x, y, y ′ ) ,
x ∈ (a, b) ,
′
a0 y(a) + b0 y (a) = c0 ,
a1 y(b) + b1 y ′ (b) = c1 .
On approximating the derivatives with divided differences, we obtain the following set of
difference equations:
yi+1 − 2yi + yi−1
h2
−3y0 + 4y1 − y2
a0 y0 + b0
2h
yN −2 − 4yN −1 + 3yN
a1 yN + b1
2h
yi+1 − yi−1
= f xi , y i ,
2h
,
i = 1, . . . , N − 1 ,
= c0 ,
= c1 .
After rearranging these, we obtain
2
1
yi+1 − yi−1
1
yi−1 − 2 yi + 2 yi+1 = f xi , yi ,
,
2
h
h
h
2h
3b0
4b0
b0
a0 −
y0 +
y1 −
y2 = c0 ,
2h
2h
2h
b1
4b1
3b1
yN −2 −
yN −1 + a1 +
yN = c1 .
2h
2h
2h
1≤i≤N −1,
This is a system of system of nonlinear equations of the form
A0 y0 + C0 y1 + B0 y2 = F0 ,
Ai yi−1 + Ci yi + Bi yi+1 = Fi ,
AN yN −2 + CN yN −1 + BN yN
i = 1, . . . , N − 1 ,
= FN ,
(115)
where
A0 = a0 − (3b0 )/(2h) , C0 = 4b0 /(2h) ,
B0 = −b0 /(2h) ,
Ci = −2/h2 ,
Bi = 1/h2 , i = 1, . . . , N − 1 ,
Ai = 1/h2 ,
AN = b1 /(2h) ,
CN = −4b1 /(2h) , BN = a1 + (3b1 )/(2h) ,
and
F0 = c0 ,
Fi = f (xi , yi , (yi+1 − yi−1 )/(2h)) ,
i = 1, . . . , N − 1 ,
FN = c1 .
Thus, (115) can be written in the compact form
M y = F(y)
where M is a tridiagonal matrix, y = (y0 , . . . , yN )T and F(y) = (F0 , . . . , FN )T . The
nonlinear system of equations
G(y) ≡ M y − F(y) = 0
69
can now be solved by Newton’s method, for example, assuming that a solution exists.
Given a starting value y(0) for the Newton iteration, subsequent iterates are computed
successively from
J(y(n) )(y(n+1) − y(n) ) = −G(y(n) ) ,
where
∂G
J(w) =
(w) =
∂y
n = 0, 1, 2, . . . ,
!
∂Gi
(w)
∂yj
(116)
0≤i,j≤N
is the Jacobi matrix of G. As Gi is independent of yj with |i − j| > 1, the matrix J(y(n) )
is tridiagonal. Consequently, each step of the Newton iteration (116) involves the solution
of a system of linear equations with a tridiagonal matrix; this may be accomplished by
using the Thomas algorithm described above.
6.3
Collocation method
Consider the boundary value problem
Ly ≡ y ′′ + p(x)y ′ + q(x)y = f (x) ,
′
a0 y(a) + b0 y (a) = c0 ,
x ∈ (a, b) ,
(117)
′
a1 y(b) + b1 y (b) = c1 ,
where a0 , a1 , b0 , b1 are real numbers such that
(a0 b1 − a1 b0 ) + a0 a1 (b − a) 6= 0 .
(118)
It can be assumed, without loss of generality, that c0 = 0 and c1 = 0; for if this is not the
case then we consider the function y˜ defined by
y˜(x) = y(x) − d(x) ,
with d(x) = αx + β where α and β are real numbers chosen so that d(x) satisfies the
(nonhomogeneous) boundary conditions at x = a and x = b stated in (117). It is a
straightforward matter to see that provided (118) holds there are unique such α and β.
The function y˜ then obeys the differential equation L˜
y = f˜, where f˜(x) = f (x) − Ld(x),
and satisfies the boundary conditions in (117) with c0 = c1 = 0.
Suppose that {ψi }N
i=1 is a set of linearly independent functions defined on the interval
[a, b] satisfying the homogeneous counterparts of the boundary conditions in (117) (i.e.
c0 = c1 = 0). We shall suppose that each function ψi is twice continuously differentiable
on [a, b].
The essence of the collocation method is to seek an approximate solution yN (x) to
(110) in the form
yN (x) =
N
X
ξi ψi (x) ,
(119)
i=1
and demand that
LyN (xj ) = f (xj ) ,
70
j = 1, . . . , N ,
(120)
at (N +1) distinct points xj , j = 1, . . . , N , referred to as the collocation points. We note
that since each of the functions ψi (x) satisfies the (homogeneous) boundary conditions at
x = a and x = b the same is true of yN (x).
Now, (119–120) yield the system of linear equations
N
X
i=1
ξi Lψi (xj ) = f (xj ) ,
j = 1, . . . , N ,
(121)
for the coefficients ξi , i = 1, . . . , N . The specific properties of the collocation method
depend on the choice of the basis functions ψi and the collocation points xj . In general
the matrix M = (Lψi (xj ))1≤i,j≤N is full. However, if each of the basis functions ψi has
compact support contained in (a, b) (for example, they are B-splines) then M is a bandmatrix, given that both ψi (xj ) = 0 and Lψi (xj ) = 0 for |i − j| > K for some integer K,
1 < K < N . If K is a fixed integer, independent of N , then the resulting system of linear
equations can be solved in O(N ) arithmetic operations.
In spectral collocation methods, the functions ψi (x) are chosen as trigonometric polynomials. Consider, for example, the simple boundary value problem
−y ′′ (x) + q(x)y(x) = f (x) ,
x ∈ (0, π) ,
y(0) = y(π) = 0 ,
where q(x) ≥ 0 for all x in [0, π]. The functions ψi (x) = sin ix, i = 1, . . . , N , satisfy the
boundary conditions and are linearly independent on [0, π]. Thus we seek an approximate
solution yN (x) in the form
yN (x) =
N
X
ξi sin ix .
i=1
Substitution of this expansion into the differential equation results in the system of linear
equations
N
X
i=1
ξi i2 + q(xj ) sin ixj = f (xj ) ,
i = 1, . . . , N ,
for ξ1 , . . . , ξN . The collocation points are usually chosen as
xj =
jπ
,
N +1
j = 1, . . . , N .
This choice is particularly convenient when q is a (nonnegative) constant, for then the
linear system
N
X
ijπ
ξi i2 + q sin
= f (xj ) ,
i = 1, . . . , N ,
N +1
i=1
can be solved for ξ1 , . . . , ξn in O(N ) arithmetic operations using a Fast Fourier Transform,
despite the fact that the matrix of the system is full.
71
Further Exercises
1. Verify that the following functions satisfy a Lipschitz condition on the respective
intervals and find the associated Lipschitz constants:
a) f (x, y) = 2yx−4 ,
b) f (x, y) =
2
e−x
tan−1 y
x ∈ [1, ∞) ;
,
x ∈ [1, ∞) ;
c) f (x, y) = 2y(1 + y 2 )−1 (1 + e−x ) ,
x ∈ (−∞, ∞) .
2. Suppose that m is a fixed positive integer. Show that the initial value problem
y ′ = y 2m/(2m+1) ,
y(0) = 0 ,
has infinitely many continuously differentiable solutions. Why does this not contradict Picard’s theorem?
3. Show that the explicit Euler method fails to approximate the solution y(x) =
(4x/5)5/4 of the initial value problem y ′ = y 1/5 , y(0) = 0. Justify your answer.
Consider the same problem with the implicit Euler method.
4. Write down the explicit Euler method for the numerical solution of the initial value
problem y ′ + 5y = xe−5x , y(0) = 0 , on the interval [0, 1] with step size h = 1/N ,
N ≥ 1. Denoting by yN the Euler approximation to y(1) at x = 1, show that
limN →∞ yN = y(1). Find an integer N0 such that
|y(1) − yN | ≤ 10−5 ,
for all N ≥ N0 .
5. Consider the initial value problem
y ′ = log log(4 + y 2 ) ,
x ∈ [0, 1] ,
y(0) = 1 ,
and the sequence {yn }N
n=0 , N ≥ 1, generated by the explicit Euler method
yn+1 = yn + h log log(4 + yn2 ) ,
n = 0, . . . , N − 1 ,
y0 = 1 ,
using the mesh points xn = nh, n = 0, . . . , N , with spacing h = 1/N . Here log
denotes the logarithm with base e.
a) Let Tn denote the truncation error of Euler’s method at x = xn for this initial
value problem. Show that |Tn | ≤ h/4.
b) Verify that
|y(xn+1 ) − yn+1 | ≤ (1 + hL)|y(xn ) − yn | + h|Tn | ,
n = 0, . . . , N − 1 ,
where L = 1/(2 log 4).
c) Find a positive integer N0 , as small as possible, such that
max |y(xn ) − yn | ≤ 10−4
0≤n≤N
whenever N ≥ N0 .
72
6. Define the truncation error Tn of the trapezium rule method
1
yn+1 = yn + h (fn+1 + fn )
2
for the numerical solution of the initial value problem y ′ = f (x, y), y(0) given, where
fn = f (xn , yn ) and h = xn+1 − xn .
By integrating by parts the integral
Z
xn+1
xn
(x − xn+1 )(x − xn )y ′′′ (x)dx ,
or otherwise, show that
1 2 ′′′
h y (ξn )
12
for some ξn in the interval (xn , xn+1 ), where y is the solution of the initial value
problem.
Tn = −
Suppose that f satisfies the Lipschitz condition
|f (x, u) − f (x, v)| ≤ L|u − v|
for all real x, u, v, where L is a positive constant independent of x, and that |y ′′′ (x)| ≤
M for some positive constant M independent of x. Show that the global error
en = y(xn ) − yn satisfies the inequality
1
1
|en+1 | ≤ |en | + hL (|en+1 | + |en |) + h2 M .
2
12
For a uniform step h satisfying hL < 2 deduce that, if y0 = y(x0 ), then
h2 M
|en | ≤
12L
1 + 12 hL
1 − 12 hL
"
!n
#
−1 .
7. Consider the following one-step method for the numerical solution of the initial value
problem y ′ = f (x, y), y(x0 ) = y0 :
1
yn+1 = yn + h(k1 + k2 ) ,
2
where
k1 = f (xn , yn ),
k2 = f (xn + h, yn + hk1 ) .
Show that the method is consistent and has truncation error
1
1
Tn = h2 fy (fx + fy f ) − (fxx + 2fxy f + fyy f 2 ) + O(h3 ) .
6
2
8. Show that for R = 1, 2 there is no R-stage Runge–Kutta method of order R + 1.
73
9. Consider the one-step method
yn+1 = yn + h(a k1 + b k2 ) ,
where
k1 = f (xn , yn ) ,
k2 = f (xn + αh, yn + βhk1 ) ,
and where a, b, α, β are real parameters. Show that there is a choice of these
parameters such that the order of the method is 2. Is there a choice of the parameters
for which the order exceeds 2?
10. Consider the one-step method
yn+1 = yn + αhf (xn , yn ) + βhf (xn + γh, yn + γhf (xn , yn )) ,
where α, β and γ are real parameters. Show that the method is consistent if and
only if α + β = 1. Show also that the order of the method cannot exceed 2.
Suppose that a second-order method of the above form is applied to the initial value
problem y ′ = −λy, y(0) = 1, where λ is a positive real number. Show that the
sequence (yn )n≥0 is bounded if and only if h ≤ λ2 . Show further that, for such λ,
1
|y(xn ) − yn | ≤ λ3 h2 xn ,
6
n≥0.
11. Derive the mid-point rule method and the Simpson rule method by integrating the
differential equation y ′ = f (x, y(x)) over suitable intervals of the real line and applying appropriate numerical integration rules.
12. Write down the general form of a linear multistep method for the numerical solution
of the initial value problem y ′ = f (x, y), y(x0 ) = y0 . What does it mean to say that
such a method is zero-stable? Explain the significance of zero-stability. What is the
truncation error of such a linear multistep method?
Determine for what values of the real parameter b the linear multistep method defined
by the formula
yn+3 + (2b − 3)(yn+2 − yn+1 ) − yn = hb(fn+2 + fn+1 )
is zero-stable. Show that there exists a value of b for which the order of the method
is 4. Show further that if this method is zero-stable then its order cannot exceed 2.
13. Consider the initial value problem y ′ = f (x, y), y(0) = y0 . Consider attempting to
solve this problem by the linear multistep method
ayn−2 + byn−1 + yn+1 = hf (xn , yn )
on the regular mesh xn = nh where a and b are constants.
a) For a certain (unique) choice of a and b, this method is consistent. Find these
values of a and b and verify that the order of accuracy is 1.
74
b) Although the method is consistent for the choice of a and b from part a), the
numerical solution it generates will not, in general, converge to the solution of
the initial value problem as h → 0, because the method is not zero-stable. Show
that the method is not zero-stable for these a and b, and describe quantitatively
what the unstable solutions will look like for small h.
14. Consider the linear two-step method
yn+2 − yn =
h
(fn+2 + 4fn+1 + fn ) .
3
Show that the method is zero-stable; show further that it is third-order accurate,
namely, Tn = O(h3 ).
15. Show that the linear three-step method
11yn+3 + 27yn+2 − 27yn+1 − 11yn = 3h[fn+3 + 9fn+2 + 9fn+1 + fn ]
is sixth order accurate. Find the roots of the first characteristic polynomial and
deduce that the method is not zero-stable.
16. Write down the general form of a linear multi-step method for the numerical solution
of the initial value problem
y ′ = f (x, y) ,
y(x0 ) = y0 ,
on the closed real interval [x0 , xN ], where f is a continuous function of its arguments
and y0 is a given real number. Define the truncation error of the method. What
does it mean to say that the method has order of accuracy p?
Given that α is a positive real number, consider the linear two-step method
yn+2 − αyn =
h
[f (xn+2 , yn+2 ) + 4f (xn+1 , yn+1 ) + f (xn , yn )] ,
3
on the mesh {xn : xn = x0 + nh, n = 1, . . . , N } of spacing h, h > 0. Determine
the set of all α such that the method is zero-stable. Find α such that the order of
accuracy is as high as possible; is the method convergent for this value of α?
17. Compute a numerical solution to the initial value problem y ′ + y = 0, y(0) = 1, on
the interval [0, 1] with h = 2−k for k = 1, 2, . . . , 10, using the linear two-step method
1
yn − yn−1 = h(fn−1 − fn−2 ) ,
2
where the missing starting value y1 is computed by the explicit Euler method. Tabulate the global error at x = 1 for k = 1, 2, . . . , 10, and comment on its rate of
decay.
18. Solve numerically the initial value problem
y′ = x − y2 ,
y(0) = 0 ,
x ∈ [0, 1] ,
with step size h = 0.01 by a fourth-order Milne-Simpson method. Use the classical
fourth-order Runge–Kutta method to compute the necessary starting values.
75
19. Find which of the following linear multistep methods for the solution of the initial
value problem y ′ = f (x, y), y(0) given, are zero-stable. For any which are zero-stable,
find limits on the value of h = xn+1 − xn for which they are absolutely stable when
applied to the equation y ′ = λy, λ < 0.
a) yn+1 − yn = hfn
b) yn+1 + yn − 2yn−1 = h(fn+1 + fn + fn−1 )
c) yn+1 − yn−1 = 13 h(fn+1 + 4fn + fn−1 )
d) yn+1 − yn = 12 h(3fn − fn1 )
e) yn+1 − yn =
1
12 h(5fn+1
+ 8fn − fn−1 )
20. Determine the order of the linear multistep method
yn+2 − (1 + a)yn+1 + yn =
1
h [(3 − a)fn+2 + (1 − 3a)fn ]
4
and investigate its zero-stability and absolute stability.
21. If σ(z) = z 2 is the second characteristic polynomial of a linear multistep method, find
a quadratic polynomial ρ(z) such that the order of the associated linear multistep
method is 2. Is this method convergent? What is its interval of absolute stability?
22. Find the interval of absolute stability for the two-step Adams–Bashforth method
1
yn+2 − yn+1 = h [3fn+1 − fn ]
2
using Schur’s criterion and the Routh–Hurwitz criterion.
23. Given that α is a positive real number, consider the linear two-step method
yn+2 − αyn =
h
[fn+2 + 4fn+1 + fn ]
3
on the mesh {xn : xn = x0 + nh, n = 0, . . . , N } of spacing h, h > 0. For values of
α such that the method is zero-stable investigate its absolute stability using Schur’s
criterion.
24. A predictor P and a corrector C are defined by their characteristic polynomials:
σ ∗ (z) = 43 (2z 3 − z 2 + 2z) ,
σ(z) = 13 (z 2 + 4z + 1) .
P : ρ∗ (z) = z 4 − 1 ,
C : ρ(z) = z 2 − 1 ,
a) Write down algorithms which use P and C in the P (EC)m E and P (EC)m
modes.
¯ and πP (EC)m (z; h)
¯ of these methb) Find the stability polynomials πP (EC)m E (z; h)
ods. Assuming that m = 1, use Schur’s criterion to calculate the associated
intervals of absolute stability.
P (EC)m E
P (EC)m
c) Express the truncation errors Tn
and Tn
of these methods in the
form O(hr ) where r = r(p∗ , p, m), with p∗ and p denoting the orders of accuracy
of P and C, respectively.
76
25. Which of the following would you consider a stiff initial value problem?
4
a) y ′ = −(105 e−10 x + 1)(y − 1), y(0) = 0 on the interval x ∈ [0, 1]. Note that the
solution can be found in closed form:
y(x) = e10(e
−104 x −1)
e−x + 1 .
b)
y1′ = −0.5y1 + 0.501y2 ,
y2′
= 0.501y1 − 0.5y2 ,
y1 (0) = 1.1 ,
y2 (0) = −0.9 ,
on the interval x ∈ [0, 103 ].
c) y′ = A(x)y, y(0) = (1, 1, 1)T , where


−1 + 100 cos 200x
100(1 − sin 200x)
0


−100(1 + sin 200x)
−(1 + 100 cos 200x)
0 .
A(x) = 
1200(cos 100x + sin 100x) 1200(cos 100x − sin 100x) −501
26. Consider the θ-method
yn+1 = yn + h [(1 − θ)fn + θfn+1 ]
for θ ∈ [0, 1].
a) Show that the method is A-stable for θ ≥ 1/2.
b) Show that the method is A(α)-stable when it is A-stable.
27. Show that the second-order backward differentiation method is A-stable. Show that
the third-order backward differentiation method is not A-stable, but that it is stiffly
stable in the sense of Definition 13.
28. Find XM > 0 as large as possible such that the system of differential equations
y1′ = −y1 + xy2
y2′ = x2 (y1 − y2 )
is dissipative in the interval [0, XM ]. Deduce that any two solutions with respective
initial conditions are then contractive on [0, XM ] in the Euclidean norm.
29. Show that the trapezium rule method is G-stable with G = 1.
30. Show that the second-order backward differentiation method and its one-leg twin
are G-stable with
"
#
5/2 −1
G=
.
−1 1/2
31. Develop a shooting method for the numerical solution of the boundary value problem
−y ′′ + yey = 1 ,
on the interval [0, 1] using:
77
y(0) = y(1) = 0 ,
a) The method of bisection;
b) The Newton-Raphson method.
Explain how your algorithm can be extended to the case of multiple shooting.
32. Construct a three-point finite difference scheme for the numerical solution of the
boundary value problem
−y ′′ + x2 y = 0 ,
y ′ (1) = 0 ,
y(0) = 1 ,
on the interval [0, 1]. Show that the resulting system of equations can be written
so that its matrix is tridiagonal. Apply the Thomas algorithm to the linear system
when the spacing between the mesh points is h = 1/10.
33. Suppose that real numbers ai , bi and ci satisfy
ai > 0 ,
bi > 0 ,
and let
ei =
ci > ai + bi ,
bi
,
ci − ai ei−1
i = 1, 2, . . . , N − 1 ,
i = 1, 2, . . . , N − 1 ,
with e0 = 0. Show by induction that 0 < ei < 1 for i = 1, 2, . . . , N − 1, and that the
conditions
ci > 0 ,
ci > |ai | + |bi | ,
i = 1, 2, . . . , N − 1 ,
are sufficient for e0 = 0 to imply that |ei | < 1 for i = 1, 2, . . . , N − 1.
How is this method used to solve the system of equations
−ai yi−1 + ci yi − bi yi+1 = fi ,
i = 1, 2, . . . , N − 1 ,
with
y0 = 0 ,
yN = 0 ?
34. Assume that the boundary value problem
y ′′ + f (x, y) = 0 ,
0<x<1,
y(0) = y(1) = 0 ,
has a unique solution with a continuous fourth derivative on the interval [0, 1]. Suppose further that a unique approximate solution can be computed satisfying the
finite difference scheme
h−2 δ2 yn + f (xn , yn ) = 0 ,
1≤n≤N −1,
y0 = yN = 0 ,
where δ2 yn ≡ yn+1 − 2yn + yn−1 , xn = nh (0 ≤ n ≤ N ), and N h = 1 for some integer
N ≥ 2.
a) Show that the truncation error of the finite difference scheme is given by
Tn =
for some ξn ∈ (xn−1 , xn+1 ).
78
1 2 (4)
h y (ξn )
12
Show further that the global error en = y(xn ) − yn satisfies
h−2 δ2 en + pn en = Tn ,
1 ≤ n≤ N −1 ,
e0 = eN = 0 ,
where pn = fy (xn , ηn ) for some ηn between y(xn ) and yn , and it is assumed
that fy (x, y) is a continuous function of x and y for x ∈ [0, 1] and y ∈ R.
b) Suppose now that fy (x, y) ≤ 0 for all x ∈ [0, 1] and y ∈ R. Let |Tn | ≤ M ,
1 ≤ n ≤ N − 1. Show that wn = 12 h2 M n(N − n) satisfies h−2 δ2 wn = −M , that
h−2 δ2 wn + pn wn ≤ −M ,
1≤n≤N −1,
and that, if vn = wn + en or vn = wn − en , then vn satisfies
h−2 δ2 vn + pn vn ≤ 0 ,
1≤n≤N −1,
v0 = vN = 0 .
c) Assume that vn has a negative minimum for some value of n between 1 and
N − 1 and show that this leads to a contradiction. Deduce that
vn ≥ 0 ,
1≤n≤N −1,
that
|en | ≤ wn ,
and that
|en | ≤
1≤n≤N −1,
1 2
h max |y (4) (x)| .
96 0≤x≤1
35. Consider the boundary value problem
−y ′′ + y ′ + y = x2 ,
y(0) = 0 ,
y(1) = 0 .
Develop a collocation method for the numerical solution of this problem on the
interval [0, 1] using the collocation points
xj =
j
,
N +1
j = 1, . . . , N .
and the basis functions ψi (x) = sin(iπx), i = 1, . . . , N . Solve the resulting system of
linear equations for increasing values of N and compare the numerical solution with
the exact solution to the problem.
79